第一部分：计算机、计算和计算

Part I: Computers, Computing, and Computation

推动人工智能 (AI) 发展的核心信念是计算机拥有制造思维所需的所有原始要素。这就是为什么人们希望通过尝试从计算系统构建智能系统来研究思维是什么以及生物系统如何使思维成为可能的原因之一。这种激励信念首先取决于对计算机是什么的理解。

The core conviction motivating artificial intelligence (AI) is that computers have all the raw ingredients they need to make minds. This is one reason why one might hope to study what minds are—and how biological systems make minds possible—by attempting to build intelligent systems out of computing systems. This motivating conviction depends, in the first instance, on an understanding of what computers are.

Haugeland 在第 2 章中对《心智设计 II》的原始介绍阐述了计算机是什么（自动形式系统）的理论，并阐明了人工智能面临的核心挑战。他的论文对本书的主题以及应该掌握的关键概念进行了出色的介绍：什么是计算机？什么是形式系统？数字和模拟有什么区别？什么是意向性？这篇论文还以原始形式包含了 Haugeland 的简洁挑战：“计算机的问题在于它们根本不在乎。”我们可以问，建立一个真正在乎的系统需要什么？

Haugeland’s original introduction to Mind Design II, given in chapter 2, lays out a theory of what computers are (automatic formal systems) and articulates the central challenges for AI. His essay provides an excellent introduction to the topics of this volume and to the key concepts one should master: What is a computer? What is a formal system? What is the difference between digital and analog? What is intentionality? This essay also contains, in primitive form, Haugeland’s pithy challenge: “The problem with computers is that they don’t give a damn.” What, we can ask, would it take to build a system that actually gives a damn?

在第 3 章中，计算机科学和人工智能的先驱 Newell 和 Simon 提出了人工智能核心的一个关键经验假设，并且因此更广泛地影响了心智计算理论：物理符号系统假设。根据他们的说法，物理符号系统是产生智能行为的必要条件。只要有资源和适当的编程，它们也足以产生智能行为。这两个主张是认知科学的基础。他们提出了以下问题：什么样的物理系统可以在不操纵符号的情况下做有趣的事情？如果认知从根本上不是符号操纵，那它又会是什么呢？

In chapter 3, Newell and Simon, pioneers of computer science and AI, advance a key empirical hypothesis at the heart of AI and, as a consequence, of computational theories of mind more generally: the Physical Symbol System Hypothesis. According to them, physical symbol systems are necessary to produce intelligent action. Given resources and appropriate programming, they are also sufficient to produce intelligent action. These two claims are foundational for much of cognitive science. They raise the questions: What kinds of physical system do interesting things without manipulating symbols? What might cognition be if it is not, at base, symbol manipulation?

大卫·马尔 (David Marr) 因其对计算解释、认知科学和神经科学之间联系的经典阐述而闻名。第 4 章中的文章摘自他的著作《视觉》( Vision) (1982)，介绍了他现在经典的三部分分析层次划分：计算层次（描述系统做什么和为什么做）、算法层次（描述表示系统和处理规则）和实现层次（描述实例化算法的硬件）。他解释了这种分工如何应用于他自己在视觉核心任务上的工作：从视网膜上的阴影模式中提取物体的形状。马尔的著名论点是，对系统的计算级理解与实例化该计算的算法和实现该计算的（生物或人工）机制无关。马尔的文章之所以重要，不仅是因为它已成为计算认知科学家的常用话题，还因为它被用来构建关于这些不同层次的解释对认知科学的相对重要性的辩论。

David Marr is well known for his classic articulation of the links between computational explanation, cognitive science, and neuroscience. The essay in chapter 4, extracted from his book, Vision (1982), introduces his now-classic tripartite division of levels of analysis: the computational level (describing what the system does and why), the algorithmic level (describing the representational system and rules for processing) and the implementational level (describing the hardware that instantiates the algorithm). He explains how this division of labor applies to his own work on a core task of vision: that of extracting shapes of objects from patterns of shading on the retina. Marr famously argues that computational-level understanding of a system is autonomous from the algorithm in which that computation is instantiated and the (biological or artificial) mechanism by which it is implemented. Marr’s essay is important not only because it has become stock in trade among computational cognitive scientists, but because it has been used to frame debates about the relative import of these different levels of explanation for cognitive science.

在本部分的最后一篇文章中，Corey Maley 讨论了重要且经常被忽视的模拟计算主题。正如他所指出的，计算的讨论通常只局限于数字计算机，而忽略了很大一部分计算可能性。数字和模拟之间的对比，以及模拟计算的独特特征，使得它们对于理解本书后面部分探讨的联结主义、动态和具体化论点尤为重要。

In the final essay of this part, Corey Maley takes up the important and often-neglected topic of analog computation. As he notes, discussions of computation often focus narrowly on digital computers, neglecting a large portion of the space of computational possibilities. The contrast between digital and analog, as well as the distinctive features of analog computation, make them especially important for understanding connectionist, dynamic, and embodied theses that are explored in later parts of the book.

2.1 观点与事物

2.1  Perspectives and things

本文的作者都不相信智慧依赖于任何非物质或超自然的东西，例如生命精神或不朽灵魂。因此，他们都是唯物主义者，至少在最低限度上认为，经过适当选择和安排的物质足以产生智慧。问题是：如何产生智慧？

None of the present authors believes that intelligence depends on anything immaterial or supernatural, such as a vital spirit or an immortal soul. Thus, they are all materialists in at least the minimal sense of supposing that matter, suitably selected and arranged, suffices for intelligence. The question is: How?

认为心灵“不过是”运动中的物质，这似乎令人难以置信。我们能想象所有这些小原子在热混沌中飞驰而过时，都在思考着深刻的思想吗？或者，如果不是一个接一个，那么也许是以万亿计的集体形式？这个难题的答案是认识到事物可以从不同的角度看待（或用不同的术语描述）——而且，当我们以不同的方式看待时，我们能够看到的东西也会不同。例如，在显微镜下看到的粗糙的磨损线编织物在商店橱窗中看到的却是一条闪亮的丝巾。在古董修复者眼中，精美的​​古老钟表在被视为废金属的价值几美分的黄铜中却如此。同样，这个想法是，从一个角度看虚空中的原子，在另一个角度看可能是一个智能系统。

It can seem incredible to suggest that mind is “nothing but” matter in motion. Are we to imagine all those little atoms thinking deep thoughts as they careen past one another in the thermal chaos? Or, if not one by one, then maybe collectively, by the zillions? The answer to this puzzle is to realize that things can be viewed from different perspectives (or described in different terms)—and, when we look differently, what we are able to see is also different. For instance, what is a coarse weave of frayed strands when viewed under a microscope is a shiny silk scarf seen in a store window. What is a marvellous old clockwork in the eyes of an antique restorer is a few cents’ worth of brass, seen as scrap metal. Likewise, so the idea goes, what is mere atoms in the void from one point of view can be an intelligent system from another.

当然，你不能随心所欲地看待任何事物——至少，不能正确地看待它。废品经销商不能把木凳看作价值几美分的黄铜，因为它不是黄铜；古物学家不能把铜猴子看作钟表，因为它不像钟表那样运转。然而，尴尬的是，这两点放在一起似乎造成了一个两难境地。根据第一点，某物是什么——粗糙还是精细，钟表还是废金属——取决于你如何看待它。但是，根据第二点，你如何正确地看待（或描述）某物取决于它是什么。人们想问，哪个先出现，是看到还是存在？

Of course, you can’t look at anything in just any way you please—at least, not and be right about it. A scrap dealer couldn’t see a wooden stool as a few cents’ worth of brass, since it isn’t brass; the antiquarian couldn’t see a brass monkey as a clockwork, since it doesn’t work like a clock. Awkwardly, however, these two points taken together seem to create a dilemma. According to the first, what something is—coarse or fine, clockwork or scrap metal—depends on how you look at it. But, according to the second, how you can rightly look at something (or describe it) depends on what it is. Which comes first, one wants to ask, seeing or being?

显然，这个问题有些问题。某物是什么和如何正确地看待它本质上并没有区别；两者都不重要，因为它们是相同的。然而，强调视角的好处是，它突显了以下问题：什么限制了如何正确地看待或描述某物（从而决定了它是什么）？这很重要，因为对于不同类型的视角或描述，答案会有所不同——正如我们的例子已经说明的那样。有时，某物是什么由其形状或形式决定（在相关的细节层面上）；有时由其组成决定；有时由其工作原理甚至只是其功能决定。其中哪些（如果有的话）可以决定某物是否（正确地看待或描述为）智能？

Clearly, there’s something wrong with that question. What something is and how it can rightly be regarded are not essentially distinct; neither comes before the other, because they are the same. The advantage of emphasizing perspective, nevertheless, is that it highlights the following question: What constrains how something can rightly be regarded or described (and thus determines what it is)? This is important, because the answer will be different for different kinds of perspective or description—as our examples already illustrate. Sometimes, what something is is determined by its shape or form (at the relevant level of detail); sometimes it is determined by what it’s made of; and sometimes by how it works or even just what it does. Which—if any—of these could determine whether something is (rightly regarded or described as) intelligent?

2.2 计算机

2.2  Computers

计算机对科学思维设计的重要性体现在两个根本不同的方面。第一个方面是很久以前图灵的灵感来源，也是最近许多其他科学家的灵感来源。但第二个方面才是真正推动人工智能发展并给人工智能带来成功希望的因素。为了理解这些各自的角色以及它们之间的区别，首先必须从根本上理解“计算机”的概念。

Computers are important to scientific mind design in two fundamentally different ways. The first is what inspired Turing long ago, and a number of other scientists much more recently. But the second is what really launched AI and gave it its first serious hope of success. In order to understand these respective roles, and how they differ, it will first be necessary to grasp the notion of ‘computer’ at an essential level.

2.3 全球化与人工智能

2.3  GOFAI

第一种方法是我所说的“传统人工智能”，即GOFAI。（有时也被称为“经典”或“符号操作”甚至“思想语言”人工智能。）GOFAI 传统研究从 50 年代中期到至少 80 年代中期占据了该领域的主导地位，原因很简单：它曾经是（现在仍然是）一种对智能机制的清晰阐述，既直观合理，又极易实现。根据这种观点，思维只是一台具有某些特殊特征的计算机，即具有可被视为明确思考或推理的内部状态和过程的计算机。为了理解 GOFAI 理念的巨大合理性和威力，我们需要了解如何正确地以这种方式看待计算机。

The first approach is what I call “good old-fashioned AI”, or GOFAI. (It is also sometimes called “classical” or “symbol-manipulation” or even “language-of-thought” Al.) Research in the GOFAI tradition dominated the field from the mid-fifties through at least the mid-eighties, and for a very good reason: it was (and still is) a well-articulated view of the mechanisms of intelligence that is both intuitively plausible and eminently realizable. According to this view, the mind just is a computer with certain special characteristics—namely, one with internal states and processes that can be regarded as explicit thinking or reasoning. In order to understand the immense plausibility and power of this GOFAI idea, we will need to see how a computer could properly be regarded in this way.

2.3.1 解释形式系统

2.3.1  Interpreted formal systems

形式系统的概念最早出现在数学中，受到算术和代数的启发。当人们解决算术或代数问题时，他们会操纵 token根据一定的规则，有点像游戏。但是这些标记与棋盘上的棋子之间存在着深刻的区别：它们有某种含义。例如，数字表示数字（具体项目或抽象项目），而算术符号表示对这些数字的运算或关系。（以这种方式表示某种含义的标记通常称为符号。）相比之下，棋子、跳棋和围棋子则没有任何含义：它们根本不是符号，而只是正式的游戏标记。

The idea of a formal system emerged first in mathematics, and was inspired by arithmetic and algebra. When people solve arithmetic or algebraic problems, they manipulate tokens according to definite rules, sort of like a game. But there is a profound difference between these tokens and, say, the pieces on a chess board: they mean something. Numerals, for instance, represent numbers (either of specified items or in the abstract), while arithmetic signs represent operations on or relationships among those numbers. (Tokens that mean something in this way are often called symbols.) Chess pieces, checkers, and go stones, by contrast, represent nothing: they are not symbols at all, but merely formal game tokens.

数学系统中的标记的操作规则与这些标记的含义密切相关。一个简单的例子可以说明这一点。假设某人正在用字母表的前 15 个字母玩正式游戏。这个游戏的规则非常严格：每个起始位置都由一串以“A”结尾的字母组成（尽管并非每个这样的字符串都是合法的）；并且，对于每个起始位置，只有一种合法的移动方式 - 即在“A”后附加一个特定的字母串（然后游戏结束）。问题是：这里发生了什么（如果有的话）？

The rules according to which the tokens in a mathematical system may be manipulated and what those tokens mean are closely related. A simple example will bring this out. Suppose someone is playing a formal game with the first fifteen letters of the alphabet. The rules of this game are very restrictive: every starting position consists of a string of letters ending in ‘A’ (though not every such string is legal); and, for each starting position, there is one and only one legal move—which is to append a particular string of letters after the ‘A’ (and then the game is over). The question is: What (if anything) is going on here?

假设你认为这些字母可能只是普通算术中熟悉的数字和符号的奇怪符号。然而，将 15 个字母转换成 15 个数字和符号的可能方法有超过一万亿种。你如何决定哪种方法（如果有的话）是“正确”的方法？表 2.1说明了这个问题。第一行给出了八个示例游戏，每个游戏都符合规则。接下来的三行分别给出了一个可能的翻译方案，并显示了根据该方案这八个示例将如何结果。

Suppose it occurs to you that the letters might be just an oddball notation for the familiar digits and signs of ordinary arithmetic. There are, however, over a trillion possible ways to translate fifteen letters into fifteen digits and signs. How could you decide which—if any—is the “right” way? The problem is illustrated in table 2.1. The first row gives eight sample games, each legal according to the rules. The next three rows each give a possible translation scheme, and show how the eight samples would come out according to that scheme.

差异是显而易见的。第一种方案呈现的样本游戏虽然由数字和算术符号组成，但看起来并不比字母更像真正的算术——它们充其量只是“算术沙拉”。第二种方案乍一看更好：至少字符串具有方程式的形状。但是，仔细检查，将它们解释为方程式，它们都会是假的——非常假。事实上，虽然符号的位置合理，但数字与第一种情况一样随机“抛出”。相比之下，第三种方案产生的字符串不仅看起来像方程式，它们就是方程式——它们都是真的。这使得第三种方案看起来更容易接受。为什么？

The differences are conspicuous. The sample games as rendered by the first scheme, though consisting of digits and arithmetic signs, look no more like real arithmetic than the letters did—they’re “arithmetic salad” at best. The second scheme, at first glance, looks better: at least the strings have the shape of equations. But, on closer examination, construed as equations, they would all be false—wildly false. In fact, though the signs are plausibly placed, the digits are just as randomly “tossed” as the first case. The third scheme, by contrast, yields strings that not only look like equations, they are equations—they’re all true. And this makes that third scheme seem much more acceptable. Why?

考虑一个相关的问题：翻译一些用迄今未知的文字书写的古代文献。显然，如果某个古怪的翻译员提出一个方案，按照这个方案，文本会变成胡言乱语（如表格中的第一个），我们不会留下深刻印象。同样明显的是，如果它们看起来像句子，但却是疯狂的句子，我们也不会留下深刻印象：不仅是错误的，而且是零散的、愚蠢的谎言，彼此之间或其他任何事情都无关。另一方面，如果某个精心设计的系统方案在其中发现了我们从其他来源了解到的关于战争、技术、自然事实或其他任何事情的详细、合理的描述，那么我们就会信服。3^但同样：为什么？

Consider a related problem: translating some ancient documents in a hitherto unknown script. Clearly, if some crank translator proposed a scheme according to which the texts came out gibberish (like the first one in the table) we would be unimpressed. Almost as obviously, we would be unimpressed if they came out looking like sentences, but loony ones: not just false, but scattered, silly falsehoods, unrelated to one another or to anything else. On the other hand, if some careful, systematic scheme finds in them detailed, sensible accounts of battles, technologies, facts of nature, or whatever, that we know about from other sources, then we will be convinced.³ But again: why?

翻译是解释的一种（见上文第 15 页）。翻译者不是说某个系统在想什么或要做什么，而是说一些标记（符号）串的含义。为了区分这两种解释，我们可以将前者称为有意解释，因为它赋予意向状态，而后者（翻译）赋予语义解释，因为它赋予意义（=语义）。

Translation is a species of interpretation (see p. 15 above). Instead of saying what some system thinks or is up to, a translator says what some strings of tokens (symbols) mean. To keep the two species distinct, we can call the former intentional interpretation, since it attributes intentional states, and the latter (translation) semantic interpretation, since it attributes meanings (= semantics).

和所有解释一样，翻译也是整体性的：不可能完全脱离上下文来解释一个简短的字符串。例如，在我们的算术示例中，合法游戏“HDJAN”碰巧在第二个方案和第三个方案中看起来都一样正确（分别为“2 × 4 = 8”和“8 − 6 = 2”）。但是，在第二个方案中，这显然只是孤立的巧合，而在第三个方案中，它是一致模式的一部分。在一组符号中寻找意义，就像在一组行为中寻找理性一样，就是在寻找某种一致、可靠的模式。

Like all interpretation, translation is holistic: it is impossible to interpret a brief string completely out of context. For instance, the legal game ‘HDJAN’ happens to come out looking just as true on the second as on the third scheme in our arithmetic example (‘2 × 4 = 8’ and ‘8 − 6 = 2’, respectively). But, in the case of the second scheme, this is obviously just an isolated coincidence, whereas, in the case of the third, it is part of a consistent pattern. Finding meaning in a body of symbols, like finding rationality in a body of behavior, is finding a certain kind of consistent, reliable pattern.

那么，是什么样的模式呢？意向性解释力求对一个系统或生物进行解释，使得它在特定情况下的想法和行为始终合理且明智。语义解释力求对一组符号进行解释，使得它们的意思（“说”）在特定情况下始终合理且明智。这就是为什么算术和古文字例子中的第三种方案是可接受的方案：它们是“理解”文本的方案，而这正是翻译所寻求的模式。我认为我们永远不会对“在特定情况下始终合理且明智”这样的短语有一个精确、明确的定义。但它肯定抓住了我们（以及图灵）所说的智能的大部分含义，无论是在行动上还是在表达上。

Well, what kind of pattern? Intentional interpretation seeks to construe a system or creature so that what it thinks and does turns out to be consistently reasonable and sensible, given its situation. Semantic interpretation seeks to construe a body of symbols so that what they mean (“say”) turns out to be consistently reasonable and sensible, given the situation. This is why the third schemes in both the arithmetic and ancient-script examples are the acceptable ones: they’re the ones that “make sense” of the texts, and that’s the kind of pattern that translation seeks. I don’t think we will ever have a precise, explicit definition of any phrase like “consistently reasonable and sensible, given the situation”. But surely it captures much of what we mean (and Turing meant) by intelligence, whether in action or in expression.

2.4 新奇的人工智能

2.4  New-fangled Al

迄今为止，这些新奇想法中最突出的——我们可以统称为NFAI ( en -fai)——属于联结主义的一般范畴。这是一个多样化且仍在快速发展的系统和提案，从表面上看，它们似乎解决了 GOFAI 的一些最明显的弱点。另一方面，联结主义系统并不那么好——至少现在还不是——能够匹敌 GOFAl 最明显的优势。（这当然表明了联合力量的可能性；但是，现在判断这样的事情是否可行还为时过早，更不用说如何做到了。）与此同时，还有其他 NFAI 想法在酝酿，既不是 GOFAI 也不是联结主义。整个领域现在比 50 年代初期更加活跃。

By far the most prominent of these new-fangled ideas—we could call them collectively NFAI (en-fai)—falls under the general rubric of connectionism. This is a diverse and still rapidly evolving bundle of systems and proposals that seem, on the face of it, to address some of GOFAI’s most glaring weaknesses. On the other hand, connectionist systems are not so good—at least not yet—at matching GOFAl’s most obvious strengths. (This suggests, of course, a possibility of joining forces; but, at this point, it’s too soon to tell whether any such thing could work, never mind how it might be done.) And, in the meantime, there are other NFAI ideas afloat, that are neither GOFAI nor connectionist. The field as a whole is in more ferment now than it has been since the earliest days, in the fifties.

2.5 心智设计缺少什么？

2.5  What’s missing from mind design?

人们普遍抱怨人工智能（无论何种类型）很少关注感觉、情感、自我、想象、情绪、意识——即内心生活的整个“现象学”。人们担心，无论机器变得多么智能，家里仍然“无人居住”。我认为这些担忧有相当多的道理，当然，有些担忧比其他担忧更甚。不过，在这里，我想简要讨论一下这种担忧的一种形式，在我看来，这种担忧比其他担忧更基本，也与狭义的认知更密切相关。

A common complaint about artificial intelligence, of whatever stripe, is that it pays scant attention to feelings, emotions, ego, imagination, moods, consciousness—the whole “phenomenology” of an inner life. No matter how smart the machines become, so the worry goes, there’s still “nobody home”. I think there is considerable merit in these misgivings, though, of course, more in some forms than in others. Here, however, I would like briefly to discuss only one form of the worry, one that strikes me as more basic than the others, and also more intimately connected with cognition narrowly conceived.

目前，人工智能领域没有一种方法能够认真对待理解——理解本身被理解为不同于知识（全部或部分）和知识的先决条件。在我看来，从这个意义上讲，只有人类才能理解任何事情——动物和人工制品都无法理解（目前还无法理解）。因此，从严格和适当的意义上讲，没有任何动物或机器真正相信或渴望任何事情——它怎么会相信它不理解的东西？——尽管，显然，在其他一些较弱的意义上，动物（至少）有很多信仰和欲望。我应该补充一点，这种信念并非基于任何原则上的障碍；据我们所知，这只是对当前情况的经验观察。

No current approach to artificial intelligence takes understanding seriously—where understanding itself is understood as distinct from knowledge (in whole or in part) and prerequisite thereto. It seems to me that, taken in this sense, only people ever understand anything—no animals and no artifacts (yet). It follows that, in a strict and proper sense, no animal or machine genuinely believes or desires anything either—How could it believe something it doesn’t understand?—though, obviously, in some other, weaker sense, animals (at least) have plenty of beliefs and desires. This conviction, I should add, is not based on any in-principle barrier; it’s just an empirical observation about what happens to be the case at the moment, so far as we can tell.

那么，系统理解某事物的意义何在？想象一下，一个系统在处理一系列对象的过程中做出或标记一系列相关区别。这些区别可以以不同的技巧反应、不同的符号结构或其他形式出现。假设，对于每个这样的区别，系统都有一个原始概念。现在我认为，系统理解其应用原始概念的对象，只要：

So, what is it for a system to understand something? Imagine a system that makes or marks a battery of related distinctions in the course of coping with some range of objects. These distinctions can show up in the form of differing skillful responses, different symbol structures, or whatever. Let’s say that, for each such distinction, the system has a proto-concept. Now I suggest that a system understands the objects to which it applies its proto-concepts insofar as:

而且，当满足这些条件时，原始概念就不仅仅是原始概念，而且是完整且真正意义上的概念。

When these conditions are met, moreover, the proto-concepts are not merely proto-concepts, but concepts in the full and proper sense.

这三个条件并非毫无关联。因为，正是在似乎不可能发生的事情面前，正确应用的问题才变得紧迫。我们可以想象系统以某种方式做出反应，我们会说：“这不可能！”然后试图找出哪里出了问题。当概念本身经常找不到任何错误时，它就需要承担责任。在这种情况下，概念结构本身必须进行修改，要么修改体现概念的辨别能力，要么修改它对什么是可能的和什么是不可能的立场，或者两者兼而有之。之后，它将拥有（或多或少）新的概念。

The three conditions are not unrelated. For, it is precisely in the face of something impossible seeming to have happened, that the question of correct application becomes urgent. We can imagine the system responding in some way that we would express by saying: “This can’t be right!” and then trying to figure out what went wrong. The responsibility for the concepts themselves emerges when, too often, it can’t find any mistake. In that event, the conceptual structure itself must be revised, either by modifying the discriminative abilities that embody the concepts, or by modifying the stand it takes on what is and isn’t possible, or both. Afterward, it will have (more or less) new concepts.

以这种方式占有并掌控其自身概念资源的系统不仅仅是在经历智能的动作，无论是进化、学习还是编程，而是为自己掌握了它们的要点。它不仅仅是做出区分或产生输出，而这些输出在我们进行最佳解释时就会变成现实。相反，这样的系统自己能够理解真理与谬误之间的区别，理解在这些方面它必须顺从世界，理解世界决定哪个是哪个——并且它^关心。我认为这就是理解。4

A system that appropriates and takes charge of its own conceptual resources in this way is not merely going through the motions of intelligence, whether evolved, learned, or programmed-in, but rather grasps the point of them for itself. It does not merely make discriminations or produce outputs that, when best interpreted by us, come out true. Rather, such a system appreciates for itself the difference between truth and falsity, appreciates that, in these, it must accede to the world, that the world determines which is which—and it cares. That, I think, is understanding.⁴

第 2 章注释

Notes for Chapter 2

1.   1.这一思想的两个部分都源于 WVO Quine 于 1960 年对意义的开创性研究。（意义是意向性的语言或符号对应物。）

2.   1.  Both parts of this idea have their roots in W.V.O. Quine’s pioneering (1960) investigations of meaning. (Meaning is the linguistic or symbolic counterpart of intentionality.)

3.   2.国际象棋选手会知道，王车易位、僵局和偷吃的规则也取决于先前的事件；因此，为了使国际象棋严格形式化，这些条件必须被编码到进一步的标记（比如说，标志）中，作为当前位置的一部分。

4.   2.  Chess players will know that the rules for castling, stalemate, and capturing en passent depend also on previous events; so, to make chess strictly formal, these conditions would have to be encoded in further tokens (markers, say) that count as part of the current position.

5.   3.密码破解（基本上就是翻译经过精心设计以使其特别困难的文本）也存在类似的观点。只有当解码后的消息始终合理、相关且真实时，密码学家才知道她成功了。

6.   3.  A similar point can be made about code-cracking (which is basically translating texts that are contrived to make that especially difficult). A cryptographer knows she has succeeded when and only when the decoded messages come out consistently sensible, relevant, and true.

7.   4. Haugeland (1997a) 的最后四章进一步探讨了这些想法。

8.   4.  These ideas are explored further in the last four chapters of Haugeland (1997a).

3.1 符号和物理符号系统

3.1  Symbols and physical symbol systems

对计算机科学知识的根本贡献之一是从相当基础的层面解释了符号是什么。这个解释是关于自然的科学命题。它是经验得出的，经过了长期而渐进的发展。

One of the fundamental contributions to knowledge of computer science has been to explain, at a rather basic level, what symbols are. This explanation is a scientific proposition about nature. It is empirically derived, with a long and gradual development.

符号是智能行为的根源，而智能行为当然是人工智能的主要主题。事实上，这是整个计算机科学的主要问题。因为所有信息都是由计算机处理以服务于目的的，我们衡量系统智能是指系统在面对任务环境的变化、困难和复杂性时，仍能实现既定目标的能力。当所完成的任务范围有限时，计算机科学在实现智能方面的这种一般投入就会被掩盖，因为那时可以准确预见环境的全部变化。当我们将计算机扩展到更全球化、更复杂和知识密集型的任务时，这一点变得更加明显——当我们试图让它们成为我们的代理，能够自行处理自然界的所有突发事件时。

Symbols lie at the root of intelligent action, which is, of course, the primary topic of artificial intelligence. For that matter, it is a primary question for all of computer science. For all information is processed by computers in the service of ends, and we measure the intelligence of a system by its ability to achieve stated ends in the face of variations, difficulties, and complexities posed by the task environment. This general investment of computer science in attaining intelligence is obscured when the tasks being accomplished are limited in scope, for then the full variations in the environment can be accurately foreseen. It becomes more obvious as we extend computers to more global, complex, and knowledge-intensive tasks—as we attempt to make them our agents, capable of handling on their own the full contingencies of the natural world.

我们对系统智能行为要求的理解是缓慢形成的。它是复合的，因为没有单一的基本事物能够解释智能的所有表现形式。没有“智能原则”，就像没有“生命原则”能够从本质上传达生命的本质一样。但是，缺乏简单的“天降神兵”并不意味着智能没有结构要求。其中一项要求是存储和操作符号的能力。为了提出这个科学问题，我们可以套用沃伦·麦卡洛克 (1961) 的一篇著名论文的标题：什么是符号，智能可以使用它，什么是智能，可以使用符号？

Our understanding of the system’s requirements for intelligent action emerges slowly. It is composite, for no single elementary thing accounts for intelligence in all its manifestations. There is no “intelligence principle”, just as there is no “vital principle” that conveys by its very nature the essence of life. But the lack of a simple deus ex machina does not imply that there are no structural requirements for intelligence. One such requirement is the ability to store and manipulate symbols. To put the scientific question, we may paraphrase the title of a famous paper by Warren McCulloch (1961): What is a symbol, that intelligence may use it, and intelligence, that it may use a symbol?

3.2 启发式搜索

3.2  Heuristic search

知道物理符号系统为智能行为提供了矩阵，并不能告诉我们它们是如何实现这一点的。我们在计算机科学中给出的第二个定性结构定律的例子解决了后一个问题，它断言符号系统通过使用启发式搜索过程来解决问题。这一概括与前一个概括一样，基于经验证据，并没有从其他前提中正式推导出来。然而，我们马上就会看到，它确实与符号系统假设有着某种逻辑上的联系，也许我们可以期待在未来的某个时候正式确定这种联系。在那一天到来之前，我们的故事必须再次成为实证探究的故事。我们将描述有关启发式搜索的已知知识，并回顾表明它如何使行动变得智能的实证研究结果。我们首先陈述这一定性结构定律，即启发式搜索假设。

Knowing that physical symbol systems provide the matrix for intelligent action does not tell us how they accomplish this. Our second example of a law of qualitative structure in computer science addresses this latter question, asserting that symbol systems solve problems by using the processes of heuristic search. This generalization, like the previous one, rests on empirical evidence, and has not been derived formally from other premises. We shall see in a moment, however, that it does have some logical connection with the symbol-system hypothesis, and perhaps we can expect to formalize the connection at some time in the future. Until that time arrives, our story must again be one of empirical inquiry. We will describe what is known about heuristic search and review the empirical findings that show how it enables action to be intelligent. We begin by stating this law of qualitative structure, the heuristic-search hypothesis.

问题搜索假设。问题的解决方案以符号结构表示。物理符号系统通过搜索来发挥其解决问题的智能——即通过生成并逐步修改符号结构，直到产生解决方案结构。

HEURISTIC-SEARCH HYPOTHESIS. The solutions to problems are represented as symbol structures. A physical symbol system exercises its intelligence in problem solving by search—that is, by generating and progressively modifying symbol structures until it produces a solution structure.

物理符号系统必须使用启发式搜索来解决问题，因为此类系统的处理资源有限；在有限数量的步骤和有限的时间间隔内，它们只能执行有限数量的进程。当然，这不是一个很强的限制，因为所有通用图灵机都受此影响。然而，我们想从更强烈的意义上来理解这种限制：我们指的是实际限制。我们可以设想一些系统，它们不受实际限制，但能够以恒定的速率并行搜索指数扩展树的节点，每个单位的深度都会增加。我们在这里不讨论这样的系统，而是讨论那些计算资源相对于所面临的情况的复杂性而稀缺的系统。这种限制不会排除任何实际任务环境中的计算机或人类符号系统。资源有限的事实使我们在大多数情况下可以将符号系统视为一个串行的、一次处理一个进程的设备。如果它在任何短时间间隔内只能完成少量处理，那么我们不妨将其视为一次只做一件事。因此，“有限资源符号系统”和“串行符号系统”实际上是同义词。如果时刻足够短，那么分配稀缺资源的问题通常可以视为调度串行机器的问题。

Physical symbol systems must use heuristic search to solve problems because such systems have limited processing resources; in a finite number of steps, and over a finite interval of time, they can execute only a finite number of processes. Of course, that is not a very strong limitation, for all universal Turing machines suffer from it. We intend the limitation, however, in a stronger sense: we mean practically limited. We can conceive of systems that are not limited in a practical way but are capable, for example, of searching in parallel the nodes of an exponentially expanding tree at a constant rate for each unit advance in depth. We will not be concerned here with such systems, but with systems whose computing resources are scarce relative to the complexity of the situations with which they are confronted. The restriction will not exclude any real symbol systems, in computer or man, in the context of real tasks. The fact of limited resources allows us, for most purposes, to view a symbol system as though it were a serial, one-process-at-a-time device. If it can accomplish only a small amount of processing in any short time interval, then we might as well regard it as doing things one at a time. Thus “limited resource symbol system” and “serial symbol system” are practically synonymous. The problem of allocating a scarce resource from moment to moment can usually be treated, if the moment is short enough, as a problem of scheduling a serial machine.

3.2.2 问题解决中的搜索

3.2.2  Search in problem solving

在人工智能研究的最初十年左右，问题解决的研究几乎与搜索过程的研究同义。从我们对问题和问题解决的特征描述中，很容易看出为什么会这样。事实上，我们可能会问，情况也可能并非如此。但在我们尝试回答这个问题之前，我们必须进一步探索搜索过程的本质，因为它在那十年的活动中显露出来。

During the first decade or so of artificial-intelligence research, the study of problem solving was almost synonymous with the study of search processes. From our characterization of problems and problem solving, it is easy to see why this was so. In fact, it might be asked whether it could be otherwise. But before we try to answer that question, we must explore further the nature of search processes as it revealed itself during that decade of activity.

从问题空间中提取信息。考虑一组符号结构，其中的一些小子集是给定问题的解决方案。进一步假设解决方案随机分布在整个集合中。这意味着不存在任何信息可以使任何搜索生成器比随机搜索表现更好。那么没有一个符号系统在解决问题时能够表现出比其他系统更高的智能（或更低的智能），尽管一个符号系统可能比另一个符号系统运气更好。

EXTRACTING INFORMATION FROM THE PROBLEM SPACE. Consider a set of symbol structures, some small subset of which are solutions to a given problem. Suppose, further, that the solutions are distributed randomly through the entire set. By this we mean that no information exists that would enable any search generator to perform better than a random search. Then no symbol system could exhibit more intelligence (or less intelligence) than any other in solving the problem, although one might experience better luck than another.

那么，出现智能的一个条件是解决方案的分布并非完全随机，符号结构空间至少表现出一定程度的秩序和模式。第二个条件是符号结构空间中的模式或多或少是可检测的。第三个条件是潜在解决方案的生成器能够根据检测到的模式做出不同的行为。问题空间中必须有信息，符号系统必须能够提取和使用这些信息。让我们首先看一个非常简单的例子，在这个例子中，智能很容易获得。

A condition, then, for the appearance of intelligence is that the distribution of solutions be not entirely random, that the space of symbol structures exhibit at least some degree of order and pattern. A second condition is that the pattern in the space of symbol structures be more or less detectable. A third condition is that the generator of potential solutions be able to behave differentially, depending on what pattern is detected. There must be information in the problem space, and the symbol system must be capable of extracting and using it. Let us look first at a very simple example, where the intelligence is easy to come by.

考虑解决一个简单代数方程的问题：

Consider the problem of solving a simple algebraic equation:

该测试将解决方案定义为任何形式为x = e的表达式，使得ae + b = ce + d。现在，人们可以使用任何可以生成数字的过程作为生成器，然后可以通过代入后一个方程来测试这些数字。我们不会将其称为智能生成器。

The test defines a solution as any expression of the form, x = e, such that ae + b = ce + d. Now, one could use as generator any process that would produce numbers which could then be tested by substituting in the latter equation. We would not call this an intelligent generator.

或者，也可以使用生成器，利用原始方程可以修改的事实（通过在两边加上或减去相同的量，或乘以或除以相同的量），而不会改变其解。但是，当然，我们可以通过将原始表达式与解的形式进行比较，并对方程进行精确的更改以使其解保持不变，同时将其转换为所需形式，从而获得更多信息来指导生成器。这样的生成器可以注意到原始方程的右边有一个不需要的cx ，从两边减去它，然后再次收集项。然后它会注意到左边有一个不需要的b并将其减去。最后，它可以通过除法去掉左边不需要的系数（ a − c ）。

Alternatively, one could use generators that would make use of the fact that the original equation can be modified—by adding or subtracting equal quantities from both sides, or multiplying or dividing both sides by the same quantity—without changing its solutions. But, of course, we can obtain even more information to guide the generator by comparing the original expression with the form of the solution, and making precisely those changes in the equation that leave its solution unchanged, while at the same time bringing it into the desired form. Such a generator could notice that there was an unwanted cx on the right-hand side of the original equation, subtract it from both sides, and collect terms again. It could then notice that there was an unwanted b on the left-hand side and subtract that. Finally, it could get rid of the unwanted coefficient (a − c) on the left-hand side by dividing.

因此，通过这个现在表现出相当高智能的程序，生成器会产生连续的符号结构，每个符号结构都是通过修改前一个符号结构获得的；修改的目的是减少输入结构形式和测试表达式形式之间的差异，同时保持解决方案的其他条件。

Thus, by this procedure, which now exhibits considerable intelligence, the generator produces successive symbol structures, each obtained by modifying the previous one; and the modifications are aimed at reducing the differences between the form of the input structure and the form of the test expression, while maintaining the other conditions for a solution.

这个简单的例子已经说明了符号系统用于智能问题解决的许多主要机制。首先，每个连续的表达式都不是独立生成的，而是通过修改先前生成的表达式而生成的。其次，修改不是随机的，而是取决于两种信息。它们取决于在整个代数问题中保持不变的信息，并且这些信息内置于生成器本身的结构中：所有表达式的修改都必须保持方程的解不变。它们还取决于每一步都会发生变化的信息：检测当前表达式和所需表达式之间仍然存在的形式差异。实际上，生成器结合了解决方案必须满足的一些测试，因此永远不会生成不满足这些测试的表达式。使用第一种信息可以保证实际上只生成所有可能表达式中的一小部分，但不会丢失该子集中的解表达式。使用第二种信息可以通过一系列近似得到所需的解决方案，采用一种简单形式的手段-目的分析来为搜索提供方向。

This simple example already illustrates many of the main mechanisms that are used by symbol systems for intelligent problem solving. First, each successive expression is not generated independently, but is produced by modifying one produced previously. Second, the modifications are not haphazard, but depend upon two kinds of information. They depend on information that is constant over this whole class of algebra problems, and that is built into the structure of the generator itself: all modifications of expressions must leave the equation’s solution unchanged. They also depend on information that changes at each step: detection of the differences in form that remain between the current expression and the desired expression. In effect, the generator incorporates some of the tests the solution must satisfy, so that expressions that don’t meet these tests will never be generated. Using the first kind of information guarantees that only a tiny subset of all possible expressions is actually generated, but without losing the solution expression from this subset. Using the second kind of information arrives at the desired solution by a succession of approximations, employing a simple form of means-ends analysis to give direction to the search.

引导搜索的信息从何而来并不神秘。我们不必效仿柏拉图，赋予符号系统一个它已经知道解决方案的先前存在。一个中等复杂的生成和测试系统可以做到这一点，而无需调用轮回。

There is no mystery where the information that guided the search came from. We need not follow Plato in endowing the symbol system with a previous existence in which it already knew the solution. A moderately sophisticated generate-and-test system did the trick without invoking reincarnation.

搜索树。简单的代数问题可能看起来是一个不寻常的、甚至是病态的搜索示例。它肯定不是反复试验搜索，因为虽然进行了几次试验，但没有错误。我们更习惯于将解决问题的搜索视为生成部分解决方案可能性的繁茂分支树，这些树在产生解决方案之前可能会增长到数千甚至数百万个分支。因此，如果生成器从它生成的每个表达式中创建B 个新分支，那么树将增长为B ^D，其中D是其深度。为代数问题生长的树具有其分支数B等于 1的特性。

SEARCH TREES. The simple algebra problem may seem an unusual, even pathological, example of search. It is certainly not trial-and-error search, for though there were a few trials, there was no error. We are more accustomed to thinking of problem-solving search as generating lushly branching trees of partial solution possibilities which may grow to thousands, or even millions, of branches, before they yield a solution. Thus, if from each expression it produces, the generator creates B new branches, then the tree will grow as B^D, where D is its depth. The tree grown for the algebra problem had the peculiarity that its branchiness, B, equaled unity.

下棋的程序通常会生成广义搜索树，在某些情况下，搜索树的分支多达一百万个或更多。虽然这个例子可以用来说明我们关于树搜索的观点，但我们应该注意，国际象棋搜索的目的不是生成建议的解决方案，而是评估（测试）它们。对游戏程序的研究方向之一是改进棋盘的表示和棋盘上的移动过程，以加快搜索速度并使搜索成为可能更大的树。当然，这个方向的理由是动态搜索越深入，最终的评估就越准确。另一方面，有充分的经验证据表明，最强大的人类棋手、大师级棋手很少探索超过一百个分支的树。实现这种经济性的原因不是搜索深度不如下棋程序，而是在每个节点上非常稀疏和有选择地分支。这只有在不导致评估恶化的情况下才有可能，方法是将更多的选择性内置到生成器本身中，以便它能够仅选择那些很可能产生有关位置的重要相关信息的分支进行生成。

Programs that play chess typically grow broad search trees, amounting in some cases to a million branches or more. Although this example will serve to illustrate our points about tree search, we should note that the purpose of search in chess is not to generate proposed solutions, but to evaluate (test) them. One line of research into gameplaying programs has been centrally concerned with improving the representation of the chess board, and the processes for making moves on it, so as to speed up search and make it possible to search larger trees. The rationale for this direction, of course, is that the deeper the dynamic search, the more accurate should be the evaluations at the end of it. On the other hand, there is good empirical evidence that the strongest human players, grandmasters, seldom explore trees of more than one hundred branches. This economy is achieved not so much by searching less deeply than do chess-playing programs, but by branching very sparsely and selectively at each node. This is only possible, without causing a deterioration of the evaluations, by having more of the selectivity built into the generator itself, so that it is able to select for generation only those branches which are very likely to yield important relevant information about the position.

这次讨论得出的结论听起来有些自相矛盾，即搜索（连续生成潜在解决方案结构）是符号系统在解决问题时运用智能的一个基本方面，但搜索量并不能衡量所展现的智能量。问题之所以成为问题，并不是因为解决问题需要进行大量搜索，而是因为如果不应用必要的智能水平，则需要进行大量搜索。当试图解决问题的符号系统对该做什么有足够的了解时，它就会直接朝着目标前进；但每当它的知识变得不足时，当它进入未知领域时，它就面临着需要进行大量搜索才能再次找到出路的威胁。

The somewhat paradoxical-sounding conclusion to which this discussion leads is that search—successive generation of potential solution structures—is a fundamental aspect of a symbol system’s exercise of intelligence in problem solving but that the amount of search is not a measure of the amount of intelligence being exhibited. What makes a problem a problem is not that a large amount of search is required for its solution, but that a large amount would be required if a requisite level of intelligence were not applied. When the symbolic system that is endeavoring to solve a problem knows enough about what to do, it simply proceeds directly towards its goal; but whenever its knowledge becomes inadequate, when it enters terra incognita, it is faced with the threat of going through large amounts of search before it finds its way again.

生成问题解决方案的每种方案中都存在搜索树呈指数级爆炸的潜力，这警告我们不要依赖计算机的蛮力——即使是最大和最快的计算机——来弥补其生成器的无知和无选择性。有些人心中仍然周期性地燃起希望，希望能够找到一台速度足够快、编程足够巧妙的计算机，通过蛮力搜索下好棋。关于国际象棋游戏的理论知识还不足以排除这种可能性。但是，对大型树中搜索管理的实证研究结果并不理想，这使得这一方向的前景远不如国际象棋首次被选为人工智能的合适任务时那么光明。我们必须将此视为国际象棋程序研究的重要实证发现之一。

The potential for the exponential explosion of the search tree that is present in every scheme for generating problem solutions warns us against depending on the brute force of computers—even the biggest and fastest computers—as a compensation for the ignorance and unselectivity of their generators. The hope is still periodically ignited in some human breasts that a computer can be found that is fast enough, and that can be programmed cleverly enough, to play good chess by brute-force search. There is nothing known in theory about the game of chess that rules out this possibility. But empirical studies on the management of search in sizable trees with only modest results make this a much less promising direction than it was when chess was first chosen as an appropriate task for artificial intelligence. We must regard this as one of the important empirical findings of research with chess programs.

智能的形式。智能的任务就是避免搜索量呈指数级爆炸式增长这一无时不在的威胁。如何实现这一目标？第一种方法是让生成器具有选择性：只生成有望成为解决方案或沿着解决方案路径前进的结构，这已在代数示例和国际象棋程序中得到说明。这样做的通常结果是降低分支率，而不是完全阻止它。最终的指数级爆炸无法避免（除非在代数示例等结构极其复杂的情形中），只能被推迟。因此，智能系统通常需要使用其他信息使用技术来补充其解决方案生成器的选择性以指导搜索。

THE FORMS OF INTELLIGENCE. The task of intelligence, then, is to avert the ever-present threat of the exponential explosion of search. How can this be accomplished? The first route, already illustrated by the algebra example and by chess programs that only generate “plausible” moves for further analysis, is to build selectivity into the generator: to generate only structures that show promise of being solutions or of being along the path toward solutions. The usual consequence of doing this is to decrease the rate of branching, not to prevent it entirely. Ultimate exponential explosion is not avoided—save in exceptionally highly structured situations like the algebra example—but only postponed. Hence, an intelligent system generally needs to supplement the selectivity of its solution generator with other information-using techniques to guide search.

二十年来，我们在各种任务环境中管理树搜索的经验已经产生了一套通用技术，这些技术是当今每位人工智能研究人员必备的装备。由于这些技术已在 Nilsson (1971) 等一般著作中描述过，因此这里可以非常简要地总结一下。

Twenty years of experience with managing tree search in a variety of task environments has produced a small kit of general techniques which is part of the equipment of every researcher in artificial intelligence today. Since these techniques have been described in general works like that of Nilsson (1971), they can be summarized very briefly here.

在串行启发式搜索中，基本问题始终是：下一步该做什么？在树搜索中，该问题又包含两个部分：(1) 我们下一步应从树中的哪个节点搜索，(2) 我们应从该节点采取什么方向？有助于回答第一个问题的信息可以解释为测量不同节点与目标之间的相对距离。最佳优先搜索要求从看起来最接近目标的节点开始进行下一步搜索。有助于回答第二个问题（即朝哪个方向搜索）的信息通常可以通过检测当前节点结构与解决方案测试所描述的目标结构之间的具体差异，并选择与减少这些特定差异相关的操作来获得，就像在代数示例中一样。这就是所谓的手段-目的分析技术，它在通用问题求解器的结构中起着核心作用。

In serial heuristic search, the basic question always is: What shall be done next? In tree search, that question, in turn, has two components: (1) From what node in the tree shall we search next, and (2) What direction shall we take from that node? Information helpful in answering the first question may be interpreted as measuring the relative distance of different nodes from the goal. Best-first search calls for searching next from the node that appears closest to the goal. Information helpful in answering the second question—in what direction to search—is often obtained, as in the algebra example, by detecting specific differences between the current nodal structure and the goal structure described by the test of a solution, and selecting actions that are relevant to reducing these particular kinds of differences. This is the technique known as means-ends analysis, which plays a central role in the structure of the General Problem Solver.

通过追溯大量问题求解程序中这两个核心思想的历史，可以清楚地证明实证研究作为人工智能研究一般思想来源的重要性：最佳优先搜索和手段-目的分析。最佳优先搜索的雏形早在 1955 年的《逻辑理论家》中就已经出现，尽管没有命名。体现手段-目的分析的通用问题求解器出现于 1957 年左右，但结合了改进的深度优先搜索而不是最佳优先搜索。出于节省内存的原因，国际象棋程序通常与深度优先搜索相结合，大约在 1958 年之后，强大的 alpha-beta 剪枝程序对其进行了补充。这些技术似乎都被重新发明了多次，直到 20 世纪 60 年代中后期，很难找到使用这些概念进行一般的、与任务无关的、理论性的问题求解讨论。他们从数学理论中获得的正式支持仍然微乎其微：一些关于使用 alpha-beta 启发式方法可以减少搜索量的定理、几个关于最短路径搜索的定理（由 Nilsson，1971 年审查），以及一些关于使用概率评估函数的最佳优先搜索的最新定理。

The importance of empirical studies as a source of general ideas in AI research can be demonstrated clearly by tracing the history, through large numbers of problem-solving programs, of these two central ideas: best-first search and means-ends analysis. Rudiments of best-first search were already present, though unnamed, in the Logic Theorist in 1955. The General Problem Solver, embodying means-ends analysis, appeared about 1957—but combined it with modified depth-first search rather than best-first search. Chess programs were generally wedded, for reasons of economy of memory, to depth-first search, supplemented after about 1958 by the powerful alpha-beta pruning procedure. Each of these techniques appears to have been reinvented a number of times, and it is hard to find general, task-independent, theoretical discussions of problem-solving in terms of these concepts until the middle or late l960’s. The amount of formal buttressing they have received from mathematical theory is still minuscule: some theorems about the reduction in search that can be secured from using the alpha-beta heuristic, a couple of theorems (reviewed by Nilsson, 1971) about shortest-path search, and some very recent theorems on best-first search with a probabilistic evaluation function.

“弱”方法和“强”方法。我们讨论的技术致力于控制指数扩张，而不是防止指数扩张。因此，它们被恰当地称为“弱方法”——当符号系统的知识或问题空间中实际包含的结构量不足以完全避免搜索时使用的方法。将高度结构化的情况（可以表述为线性规划问题）与组合问题中结构性较差的情况，如旅行商问题或调度问题。（“结构性较差”在这里指的是问题空间结构的相关理论不足或不存在。）

“WEAK” AND “STRONG” METHODS. The techniques we have been discussing are dedicated to the control of exponential expansion rather than its prevention. For this reason, they have been properly called “weak methods”—methods to be used when the symbol system’s knowledge or the amount of structure actually contained in the problem space are inadequate to permit search to be avoided entirely. It is instructive to contrast a highly-structured situation, which can be formulated, say, as a linear-programming problem, with the less-structured situations of combinatorial problems like the traveling-salesman problem or scheduling problems. (“Less structured” here refers to the insufficiency or nonexistence of relevant theory about the structure of the problem space.)

在解决线性规划问题时，可能需要进行大量计算，但搜索不会出现分支。每一步都是通往解决方案的一步。在解决组合问题或证明定理时，树搜索几乎是不可避免的，成功取决于我们所描述的那种启发式搜索方法。

In solving linear-programming problems, a substantial amount of computation may be required, but the search does not branch. Every step is a step along the way to a solution. In solving combinatorial problems or in proving theorems, tree search can seldom be avoided, and success depends on heuristic search methods of the sort we have been describing.

并非所有人工智能问题解决研究都遵循了我们概述的路径。定理证明系统的研究提供了一个略有不同的观点的例子。在这里，从数学和逻辑中引入的思想对研究方向产生了强烈的影响。例如，当无法证明完整性属性时，人们会抵制使用启发式方法（这有点讽刺，因为大多数有趣的数学系统都是不可判定的）。由于很少能证明最佳优先搜索启发式方法或多种选择性生成器的完整性，因此这一要求的影响相当令人沮丧。当定理证明程序因其搜索树的组合爆炸而不断无能为力时，人们开始考虑选择性启发式方法，在许多情况下，选择性启发式方法被证明是一般问题解决程序中使用的启发式方法的类似物。例如，支持集启发式方法是一种逆向工作形式，适用于解析定理证明环境。

Not all streams of AI problem-solving research have followed the path we have been outlining. An example of a somewhat different point is provided by the work on theorem-proving systems. Here, ideas imported from mathematics and logic have had a strong influence on the direction of inquiry. For example, the use of heuristics was resisted when properties of completeness could not be proved (a bit ironic, since most interesting mathematical systems are known to be undecidable). Since completeness can seldom be proved for best-first search heuristics, or for many kinds of selective generators, the effect of this requirement was rather inhibiting. When theorem-proving programs were continually incapacitated by the combinatorial explosion of their search trees, thought began to be given to selective heuristics, which in many cases proved to be analogues of heuristics used in general problem-solving programs. The set-of-support heuristic, for example, is a form of working backward, adapted to the resolution theorem-proving environment.

经验总结。我们现在已经描述了定性结构第二定律的工作原理，该定律断言物理符号系统通过启发式搜索解决问题。除此之外，我们还研究了启发式搜索的一些次要特征，特别是它始终面临的搜索树指数爆炸的威胁，以及它用来避免这种威胁的一些手段。对于启发式搜索作为一种解决问题的机制有多有效，人们的看法不一——这些看法取决于考虑的任务领域和采用的充分性标准。将期望水平设定得低可以保证成功，而将期望水平设定得高则会导致失败。证据可以总结如下：很少有程序能以“专家”专业水平解决问题。塞缪尔的检查程序和费根鲍姆和莱德伯格的 DENDRAL 可能是最著名的例外，但人们也可以指出许多启发式搜索程序，用于调度和整数规划等运筹学问题领域。在许多领域，程序的表现都与业余爱好者相当：国际象棋、某些定理证明领域、多种游戏和谜题。具有复杂感知“前端”的程序还远未达到人类的水平：视觉场景识别器、语音理解器、必须在真实空间和时间中移动的机器人。尽管如此，这些困难任务已经取得了令人瞩目的进展，并积累了大量经验。

A SUMMARY OF THE EXPERIENCE. We have now described the workings of our second law of qualitative structure, which asserts that physical symbol systems solve problems by means of heuristic search. Beyond that, we have examined some subsidiary characteristics of heuristic search, in particular the threat that it always faces of exponential explosion of the search tree, and some of the means it uses to avert that threat. Opinions differ as to how effective heuristic search has been as a problem-solving mechanism—the opinions depending on what task domains are considered and what criterion of adequacy is adopted. Success can be guaranteed by setting aspiration levels low—or failure by setting them high. The evidence might be summed up about as follows: Few programs are solving problems at “expert” professional levels. Samuel’s checker program and Feigenbaum and Lederberg’s DENDRAL are perhaps the best-known exceptions, but one could point also to a number of heuristic search programs for such operations-research problem domains as scheduling and integer programming. In a number of domains, programs perform at the level of competent amateurs: chess, some theorem-proving domains, many kinds of games and puzzles. Human levels have not yet been nearly reached by programs that have a complex perceptual “front end”: visual-scene recognizers, speech understanders, robots that have to maneuver in real space and time. Nevertheless, impressive progress has been made, and a large body of experience assembled about these difficult tasks.

对于出现的特定绩效模式，我们并没有深刻的理论解释。然而，从实证的角度，我们可以得出两个结论。首先，从人类专家在国际象棋等任务中的表现来看，任何能够匹敌这种表现的系统都必须在其内存中存储大量语义信息。其次，人类在感知成分较大的任务中的优势部分可以归因于人眼和耳朵内置的特殊并行处理结构。

We do not have deep theoretical explanations for the particular pattern of performance that has emerged. On empirical grounds, however, we might draw two conclusions. First, from what has been learned about human expert performance in tasks like chess, it is likely that any system capable of matching that performance will have to have access, in its memories, to very large stores of semantic information. Second, some part of the human superiority in tasks with a large perceptual component can be attributed to the special-purpose built-in parallel-processing structure of the human eye and ear.

无论如何，性能的质量必然取决于问题领域和用于解决问题的符号系统的特征。对于我们感兴趣的大多数现实领域，迄今为止，领域结构还不够简单，无法产生关于复杂性的定理，或者无法告诉我们（除了经验之外）现实世界的问题与我们的符号系统解决这些问题的能力之间的关系有多大。这种情况可能会改变，但在改变之前，我们必须依靠经验探索，使用我们知道如何构建的最佳问题求解器，作为了解问题难度大小和特征的主要知识来源。即使在线性规划等高度结构化的领域，理论在加强最强大的解决方案算法所依赖的启发式方法方面也比在提供复杂性的深入分析方面更有用。

In any case, the quality of performance must necessarily depend on the characteristics both of the problem domains and of the symbol systems used to tackle them. For most real-life domains in which we are interested, the domain structure has so far not proved sufficiently simple to yield theorems about complexity, or to tell us, other than empirically, how large real-world problems are in relation to the abilities of our symbol systems to solve them. That situation may change, but until it does, we must rely upon empirical explorations, using the best problem solvers we know how to build, as a principal source of knowledge about the magnitude and characteristics of problem difficulty. Even in highly structured areas like linear programming, theory has been much more useful in strengthening the heuristics that underlie the most powerful solution algorithms than in providing a deep analysis of complexity.

第 3 章注释

Notes for Chapter 3

1.   1.编者注：术语“指示”和“解释”的含义，以及“符号”的含义，是计算机科学所特有的；它们只涉及计算机内部发生的关系和过程。  相反，在语言学和哲学中，这些术语通常用智能系统（或其内部的东西）与其环境之间的关系来解释。本卷中的大多数文章都使用后一种含义。

2.   1.  Editor’s note: These senses of the terms ‘designation’ and ‘interpretation’, and hence also of ‘symbol’, are specific to computer science; they concern only relationships and processes that occur within a computer. In linguistics and philosophy, by contrast, these terms would usually be explained in terms of relationships between an intelligent system (or what’s inside of it) and its environment. Most of the essays in the present volume use the terms in this latter sense.

理解复杂的信息处理系统（视觉，第 1.2 节）

Understanding Complex Information-Processing Systems (Vision, Section 1.2)

几乎从来都不能将任何类型的复杂系统理解为对其基本成分属性的简单推断。例如，考虑瓶子中的气体。热力学效应（温度、压力、密度以及这些因素之间的关系）的描述不是通过使用大量方程式（每个相关粒子都有一个方程式）来制定的。这些效应是在它们自己的层面上描述的，即在大量粒子的层面上；努力是为了表明微观和宏观描述在原则上是相互一致的。如果人们希望全面理解像神经系统、发育中的胚胎、一组代谢途径、一瓶气体甚至大型计算机程序这样复杂的系统，那么人们必须准备好考虑不同描述层次上的不同类型的解释，这些解释至少在原则上是联系在一起的，形成一个有凝聚力的整体，即使将各个层次完全详细地联系起来是不切实际的。对于解决信息处理问题的系统的具体情况，还有过程和表示的双重线索，这两个想法都需要讨论。

Almost never can a complex system of any kind be understood as a simple extrapolation from the properties of its elementary components. Consider, for example, some gas in a bottle. A description of thermodynamic effects—temperature, pressure, density and the relationships among these factors—is not formulated by using a large set of equations, one for each of the particles involved. Such effects are described at their own level, that of an enormous collection of particles; the effort is to show that in principle the microscopic and macroscopic descriptions are consistent with one another. If one hopes to achieve a full understanding of a system as complicated as a nervous system, a developing embryo, a set of metabolic pathways, a bottle of gas, or even a large computer program, then one must be prepared to contemplate different kinds of explanation at different levels of description that are linked, at least in principle, into a cohesive whole, even if linking the levels in complete detail is impractical. For the specific case of a system that solves an information-processing problem, there are in addition the twin strands of process and representation, and both these ideas need some discussion.

三个层次

The Three Levels

我们可以按照图 4.1中所示的方式总结我们的讨论，该图说明了在完全理解信息处理设备之前，必须理解它的不同层次。在一个极端，即最高层次，是设备的抽象计算理论，其中设备的性能被描述为从一种信息到另一种信息的映射，这种映射的抽象属性被精确定义，并且它对于当前任务的适当性和充分性得到了证明。在中心是输入和输出表示的选择，以及将一个表示转换为另一个表示所使用的算法。而在另一个极端是算法和表示的物理实现细节——可以说是详细的计算机架构。这三个层次是耦合的，但只是松散的。例如，算法的选择受算法必须执行的操作和必须运行的硬件的影响。但每个层次都有广泛的选择，每个层次的阐述都涉及与其他两个层次相当独立的问题。

We can summarize our discussion in something like the manner shown in Figure 4.1, which illustrates the different levels at which an information-processing device must be understood before one can be said to have understood it completely. At one extreme, the top level, is the abstract computational theory of the device, in which the performance of the device is characterized as a mapping from one kind of information to another, the abstract properties of this mapping are defined precisely, and its appropriateness and adequacy for the task at hand are demonstrated. In the center is the choice of representation for the input and output and the algorithm to be used to transform one into the other. And at the other extreme are the details of how the algorithm and representation are realized physically—the detailed computer architecture, so to speak. These three levels are coupled, but only loosely. The choice of an algorithm is influenced for example, by what it has to do and by the hardware in which it must run. But there is a wide choice available at each level, and the explication of each level involves issues that are rather independent of the other two.

三个描述层次中的每一个都将在最终理解感知信息处理中占有一席之地，当然它们在逻辑上和因果上是相关的。但需要注意的一点是，由于这三个层次之间的联系相当松散，一些现象只能在其中的一个或两个层次上得到解释。这意味着，例如，对一些心理物理观察的正确解释必须在适当的层次上形成。在试图将心理物理问题与生理学联系起来时，人们常常对应该在哪个层次上解决问题感到困惑。例如，有些问题主要与视觉的物理机制有关——如后像（例如，盯着灯泡后看到的后像）或任何颜色都可以通过三原色的适当混合来匹配（这主要是由于我们人类有三种类型的视锥细胞）。另一方面，内克尔立方体（图 4.2）的模糊性似乎需要不同类型的解释。可以肯定的是，其感知逆转的部分解释必定与大脑内部某处的双稳态神经网络（即具有两个不同稳定状态的神经网络）有关，但很少有人会对没有提到存在两种不同但完全合理的二维图像三维解释的解释感到满意。

Each of the three levels of description will have its place in the eventual understanding of perceptual information processing, and of course they are logically and causally related. But an important point to note is that since the three levels are only rather loosely related, some phenomena may be explained at only one or two of them. This means, for example, that a correct explanation of some psychophysical observation must be formulated at the appropriate level. In attempts to relate psychophysical problems to physiology, too often there is confusion about the level at which problems should be addressed. For instance, some are related mainly to the physical mechanisms of vision—such as afterimages (for example, the one you see after staring at a light bulb) or such as the fact that any color can be matched by a suitable mixture of the three primaries (a consequence principally of the fact that we humans have three types of cones). On the other hand, the ambiguity of the Necker cube (Figure 4.2) seems to demand a different kind of explanation. To be sure, part of the explanation of its perceptual reversal must have to do with a bistable neural network (that is, one with two distinct stable states) somewhere inside the brain, but few would feel satisfied by an account that failed to mention the existence of two different but perfectly plausible three-dimensional interpretations of this two-dimensional image.

对于某些现象，所需的解释类型相当明显。例如，神经解剖学显然主要与第三层级有关，即计算的物理实现。突触机制、动作电位、抑制相互作用等也是如此。神经生理学也主要与这一层级相关……但人们在从神经生理学发现中推断时必须极其谨慎所使用的算法和表示，特别是直到人们清楚地知道需要表示什么信息以及需要实施什么过程。

For some phenomena, the type of explanation required is fairly obvious. Neuroanatomy, for example, is clearly tied principally to the third level, the physical realization of the computation. The same holds for synaptic mechanisms, action potentials, inhibitory interactions, and so forth. Neurophysiology, too, is related mostly to this level…But one has to exercise extreme caution in making inferences from neurophysiological findings about the algorithms and representations being used, particularly until one has a clear idea about what information needs to be represented and what processes need to be implemented.

另一方面，心理物理学与算法和表示的水平更直接相关。不同的算法在性能达到极限或缺乏关键信息时，往往会以截然不同的方式失败。正如我们将看到的，主要是心理物理学证据向 Poggio 和我证明了我们的第一个立体匹配算法（Marr 和 Poggio，1976 年）不是大脑使用的算法，而我们的第二个算法（Marr 和 Poggio，1979 年）大致是大脑使用的算法的最佳证据也来自心理物理学。当然，在这两种情况下，底层计算理论保持不变，只是算法不同。

Psychophysics, on the other hand, is related more directly to the level of algorithm and representation. Different algorithms tend to fail in radically different ways as they are pushed to the limits of their performance or are deprived of critical information. As we shall see, primarily psychophysical evidence proved to Poggio and myself that our first stereo-matching algorithm (Marr and Poggio, 1976) was not the one that is used by the brain, and the best evidence that our second algorithm (Marr and Poggio, 1979) is roughly the one that is used also comes from psychophysics. Of course, the underlying computational theory remained the same in both cases, only the algorithms were different.

心理物理学也有助于确定表征的性质。Roger Shepard (1975)、Eleanor Rosch (1978) 或 Elizabeth Warrington (1975) 的工作在这方面提供了一些有趣的提示。更具体地说，Stevens (1979) 从心理物理实验中论证道，表面方向由倾斜和倾角的坐标表示，而不是（例如）更传统的梯度空间 ( p, q )。他还从受试者在广泛方向上判断表面方向时所犯错误大小的一致性推断出，用于倾斜和倾角的表征量是纯角​​度，而不是它们的余弦、正弦或正切。

Psychophysics can also help to determine the nature of a representation. The work of Roger Shepard (1975), Eleanor Rosch (1978), or Elizabeth Warrington (1975) provides some interesting hints in this direction. More specifically, Stevens (1979) argued from psychophysical experiments that surface orientation is represented by the coordinates of slant and tilt, rather than (for example) the more traditional (p, q) of gradient space. He also deduced from the uniformity of the size of errors made by subjects judging surface orientation over a wide range of orientations that the representational quantities used for slant and tilt are pure angles and not, for example, their cosines, sines, or tangents.

更一般地讲，如果能清楚地记住不同现象需要在不同层次上解释这一观点，那么它常常有助于评估不时提出的各种反对意见的有效性。例如，人们最喜欢的一种观点是，大脑与计算机完全不同，因为一个是并行的，另一个是串行的。当然，答案是，串行和并行之间的区别是算法层面的区别；它根本不是根本的区别——任何并行编程的东西都可以串行重写（尽管反之亦然）。因此，这种区别并不能成为争论大脑与计算机的运作方式如此不同以至于计算机无法被编程来执行相同任务的理由。

More generally, if the idea that different phenomena need to be explained at different levels is kept clearly in mind, it often helps in the assessment of the validity of the different kinds of objections that are raised from time to time. For example, one favorite is that the brain is quite different from a computer because one is parallel and the other serial. The answer to this, of course, is that the distinction between serial and parallel is a distinction at the level of algorithm; it is not fundamental at all—anything programmed in parallel can be rewritten serially (though not necessarily vice versa). The distinction, therefore, provides no grounds for arguing that the brain operates so differently from a computer that a computer could not be programmed to perform the same tasks.

概要（愿景，第 6 章）

Synopsis (Vision, Chapter 6)

我们对这种新的视觉计算方法的调查现已完成。尽管该说明中存在许多空白，但我希望它足够可靠，能够建立关于该主题的坚定观点，并促使读者开始判断其价值。在这一简短的章节中，我将对整个方法进行非常广泛的介绍，探究其最重要的一般特征及其相互关系，并尝试谈谈这种方法所暗示的研究风格。将讨论分为四个要点很方便。

Our survey of this new, computational approach to vision is now complete. Although there are many gaps in the account, I hope that it is solid enough to establish a firm point of view about the subject and to prompt the reader to begin to judge its value. In this brief chapter, I shall take a very broad view of the whole approach, inquiring into its most important general features and how they relate to one another, and trying to say something about the style of research that this approach implies. It is convenient to divide the discussion into four main points.

第一点是我们在整个叙述中都遇到过的——不同层次解释的概念。这种方法的核心原则是，要理解视觉是什么以及它是如何工作的，只在一个层次上的理解是不够的。仅仅能够描述单个细胞的反应是不够的，仅仅能够局部预测心理物理实验的结果也是不够的。甚至能够编写出大致按照预期方式运行的计算机程序也是不够的。人们必须同时做所有这些事情，同时也要非常清楚我称之为计算理论层次的额外解释层次。认识到这个层次的存在和重要性是这种方法最重要的方面之一。认识到这一点后，人们可以明确地制定三个层次的解释（计算理论、算法和实现），然后就会清楚地知道这些不同的层次与可以进行的不同类型的经验观察和理论分析之间的关系。我特别强调计算理论层面，并不是因为我认为它本身比其他两个层面更重要——这种方法的真正威力在于整合所有三个攻击层面——而是因为这是一个以前从未被认识和付诸行动的解释层面。因此，它可能是该领域新手最难理解的概念之一，仅凭这个原因，它的重要性在任何一本入门书中都不应被低估，本书就是如此。

The first point is one that we have met throughout the account—the notion of different levels of explanation. The central tenet of the approach is that to understand what vision is and how it works, an understanding at only one level is insufficient. It is not enough to be able to describe the responses of single cells, nor is it enough to be able to predict locally the results of psychophysical experiments. Nor it is enough even to be able to write computer programs that perform approximately in the desired way. One has to do all these things at once and also be very aware of the additional level of explanation that I have called the level of computational theory. The recognition of the existence and importance of this level is one of the most important aspects of this approach. Having recognized this, one can formulate the three levels of explanation explicitly (computational theory, algorithm, and implementation), and it then becomes clear how these different levels are related to the different types of empirical observation and theoretical analysis that can be conducted. I have laid particular stress on the level of computational theory not because I regard it as inherently more important than the other two levels—the real power of the approach lies in the integration of all three levels of attack—but because it is a level of explanation that has not previously been recognized and acted upon. It is therefore probably one of the most difficult ideas for newcomers to the field to grasp, and for this reason alone its importance should not be understated in any introductory book, such as this is intended to be.

第二个要点是，从信息处理的角度看，我们能够为视觉过程制定一个相当清晰的整体框架。这个框架基于这样的观点：视觉中的关键问题围绕着所用表征的性质——即在视觉过程中明确表达的世界的特定特征——以及恢复这些特征、创建和维护表征并最终读取表征的过程的性质。通过分析空间通过对视觉问题各个方面进行分析，我们得出了视觉信息处理的总体框架，该框架基于三个主要表示：（1）原始草图，涉及二维图像的明确属性，从强度变化的数量和配置到局部图像几何的原始表示，更复杂的是，包括底层反射分布中存在的任何高阶结构的分层描述；（2）-D草图，以观察者为中心，表示可见表面的深度和方向，包括这些量的不连续轮廓；（3）3-D模型表示，其重要特征是它的坐标系以物体为中心，包括体积基元（明确表示物体所占据的空间的组织，而不仅仅是其可见表面），包括各种大小的基元，以模块化、分层的方式组织起来。

The second main point is that by taking an information-processing point of view, we have been able to formulate a rather clear overall framework for the process of vision. This framework is based on the idea that the critical issues in vision revolve around the nature of the representations used—that is, the particular characteristics of the world that are made explicit during vision—and the nature of the processes that recover these characteristics, create and maintain the representations, and eventually read them. By analyzing the spatial aspects of the problem of vision, we arrived at an overall framework for visual information processing that hinges on three principal representations: (1) the primal sketch, which is concerned with making explicit properties of the two-dimensional image, ranging from the amount and disposition of the intensity changes there to primitive representations of the local image geometry and including at the more sophisticated end a hierarchical description of any higher-order structure present in the underlying reflectance distributions; (2) the -D sketch, which is a viewer-centered representation of the depth and orientation of the visible surfaces and includes contours of discontinuities in these quantities; and (3) the 3-D model representation, whose important features are that its coordinate system is object centered, that it includes volumetric primitives (which make explicit the organization of the space occupied by an object and not just its visible surfaces), and that primitives of various size are included, arranged in a modular, hierarchical organization.

第三个要点涉及从场景图像中恢复场景物理特征各个方面的过程的研究。制定此类过程的计算理论的关键是发现对世界行为方式的有效约束，这些约束可提供足够的附加信息以恢复所需的特征。这种分析的力量在于，发现有效、足够普遍的约束可以得出与视觉相关的结论，这些结论与其他科学分支的结论一样具有持久性。

The third main point concerns the study of processes for recovering the various aspects of the physical characteristics of a scene from images of it. The critical act in formulating computational theories for such processes is the discovery of valid constraints on the way the world behaves that provide sufficient additional information to allow recovery of the desired characteristic. The power of this type of analysis resides in the fact that the discovery of valid, sufficiently universal constraints leads to conclusions about vision that have the same permanence as conclusions in other branches of science.

此外，一旦制定了某个过程的计算理论，就可以设计出实现该过程的算法，并将其性能与人类视觉处理器的性能进行比较。这允许两种结果。首先，如果性能基本相同，我们有充分的证据表明底层计算理论的约束是有效的，并且可能隐含在人类处理器中；其次，如果某个过程与人类的表现相匹配，那么它可能足够强大，可以成为通用视觉机器的一部分。

Furthermore, once a computational theory for a process has been formulated, algorithms for implementing it may be designed, and their performance compared with that of the human visual processor. This allows two kinds of results. First, if performance is essentially identical, we have good evidence that the constraints of the underlying computational theory are valid and may be implicit in the human processor; second, if a process matches human performance, it is probably sufficiently powerful to form part of a general purpose vision machine.

最后一点涉及这类方法的方法论或风格，它涉及两个主要观察结果。首先，图 4.3中明确列出的表征和过程之间的二元性通常有助于思考在研究特定问题时如何最好地进行。在表征和过程的研究中，日常经验或相当普遍的心理物理学甚至神经生理学发现往往会引发一般问题。这种一般性观察结果通常可以导致特定过程或表征理论的形成，可以对其具体示例进行编程或对其进行详细的心理物理测试。一旦我们对这一层面的过程或表征的正确性有足够的信心，我们就可以探究其详细的实现，这涉及到神经生理学和神经解剖学的终极和非常困难的问题。

The final point concerns the methodology or style of this type of approach, and it involves two main observations. First, the duality between representations and processes, which is set out explicitly in Figure 4.3, often provides a useful aid to thinking how best to proceed when studying a particular problem. In the study both of representations and of processes, general problems are often suggested by everyday experience or by psychophysical or even neurophysiological findings of a quite general nature. Such general observations can often lead to the formulation of a particular process or representational theory, specific examples of which can be programmed or subjected to detailed psychophysical testing. Once we have sufficient confidence in the correctness of the process or representation at this level, we can inquire about its detailed implementation, which involves the ultimate and very difficult problems of neurophysiology and neuroanatomy.

第二个观察结果是，这类研究没有真正的秘诀——尽管我有时暗示有——就像在任何其他科学分支中，没有发现事物的直接程序一样。事实上，乐趣的一部分在于我们永远不知道下一个关键点会从何而来——日常经验、神经缺陷报告、三维几何定理、超敏反应中的心理物理发现、神经生理观察或表征问题的仔细分析。所有这些类型的信息在建立我所描述的框架方面都发挥了重要作用，它们可能将继续以一种有趣且不可预测的方式为其进步做出贡献。我只希望这些观察可以说服我的一些读者加入我们的冒险之旅，并帮助完成解开人类视觉感知之谜这一漫长而有意义的任务。

The second observation is that there is no real recipe for this type of research—even though I have sometimes suggested that there is—any more than there is a straightforward procedure for discovering things in any other branch of science. Indeed, part of the fun is that we never really know where the next key is going to come from—a piece of daily experience, the report of a neurological deficit, a theorem about three-dimensional geometry, a psychophysical finding in hyperacuity, a neurophysiological observation, or the careful analysis of a representational problem. All these kinds of information have played important roles in establishing the framework that I have described, and they will presumably continue to contribute to its advancement in an interesting and unpredictable way. I hope only that these observations may persuade some of my readers to join in the adventures we have had and to help in the long but rewarding task of unraveling the mysteries of human visual perception.

5.1 计算机器简史

5.1  A Brief History of Computational Machines

由于本章的大部分内容都致力于纠正我认为目前对计算类型和表征类型之间基本划分理解上的错误，因此花点时间看看我们是如何走到这一步的会很有帮助。然而有趣的是，关于计算历史的著作并不像人们想象的那么多。Mahoney (2011) 指出，技术史学家对数字计算机历史的著述相对较少。1^著述的作者主要是计算机科学家和数学家，这给计算历史学家和哲学家带来了问题：

Because much of this chapter is devoted to correcting what I take to be an error in how fundamental divisions between computational and representational types are currently understood, it is helpful to take a moment to see how we got here. Interestingly, however, there is not as much written on the history of computation as one might expect. Mahoney (2011) notes that historians of technology have written relatively little on the history of digital computing machines.¹ The authors of what has been written are largely computer scientists and mathematicians, which presents problems to the historian and philosopher of computation:

这个问题在涉及模拟计算机时尤其严重，因为关于模拟计算机的描述甚至更少（Mindell，2002 年）。Nyce（1996 年，3）很好地阐述了这一点：“由于数字计算机和计算非常成功，它们影响了我们如何看待计算机作为机器以及计算作为一个过程——以至于今天很难重建模拟计算的全部内容。”完整的历史将不得不等待另一个时代（和它自己的书）；现在，我们可以勾勒出这段历史，足以让我们感受到未来会发生什么。

This problem is particularly acute when it comes to analog computing machines, where even less has been written (Mindell, 2002). Nyce (1996, 3) puts the point well: “Because digital computers and computation have been so successful, they have influenced how we think about both computers as machines and computation as a process—so much so, it is difficult today to reconstruct what analog computing was all about.” A complete history will have to wait for another time (and its own book); for now, we can sketch that history just enough to get a feel for what is to come.

在十九世纪末和二十世纪初，各种各样的计算机器被用于科学、工程和工业领域。其中包括加法机、各种类型的计算尺和机械模拟计算机。一般来说，这种大杂烩机制没有经过任何系统的理论研究：不同的问题需要不同的解决方案，大多数机器都是为解决特定类型的任务而制造的。例如，机械加法器的工作原理与收银机非常相似，允许人类计算机²可以执行基本的算术计算。制表机使操作员能够分析人口普查数据。在海军战舰上，机械火控计算机决定如何定位舰炮以准确击中移动目标（Mindell，2002 年）。

In the late nineteenth and early twentieth century, a wide variety of computing machines were used in scientific, engineering, and industrial contexts. These included adding machines, slide rules of various types, and mechanical analog computers. In general, this hodgepodge of mechanisms was not subject to any systematic theoretical study: different problems required different solutions, and most machines were built to solve a particular kind of task. For example, mechanical adders worked much like cash registers, allowing human computers² to perform basic arithmetical calculations. Tabulating machines allowed operators to analyze census data. On naval warships, mechanical fire-control computing machines determined how to position the ship’s guns to accurately hit moving targets (Mindell, 2002).

缺乏系统分析的一个例外是微分分析仪的研究，它是麻省理工学院的 Vannevar Bush 开发的一种机械模拟计算机。与许多其他为单一特殊目的而设计的计算机不同，微分分析仪可以重新配置以计算大量不同类型数学问题的解 (Bush, 1931)。这些机械模拟计算机主要（但不完全）用于解决涉及微分方程的问题。许多工程和科学问题都可以用微分方程组来分析，但这些系统中的大部分没有解析解。因此，它们只能通过繁琐的数值计算或某种机械模拟来“解决”。事实证明，微分分析仪是研究此类问题的一种有用的新方法。

One exception to this lack of systematic analysis was the work done on the differential analyzer, a kind of mechanical analog computer developed by Vannevar Bush at MIT. Unlike many other computing machines that were created for a single, special purpose, the differential analyzer could be reconfigured to compute solutions to a large number of different kinds of mathematical problems (Bush, 1931). These mechanical analog computers were mostly (but not exclusively) used for solving problems that involved differential equations. Many engineering and scientific problems are amenable to analysis in terms of systems of differential equations, but large classes of these systems have no analytic solution. Thus, they can only be “solved” by tedious numerical calculation, or by some kind of mechanical simulation. The differential analyzer proved a useful new way to study problems of this sort.

纯机械模拟计算机最终让位于机电和电子计算机。机械模拟计算机中的变量将通过轴的左右位移、齿轮的角度或滚筒或滚筒的总旋转次数等量来表示：换句话说，要解决的问题量由计算机中的物理量表示，而这些量的变化则通过物理移动这些部件来表示。机电和电子计算机使用电量（例如电压和电阻）来表示变量的数量。并且，与机械模拟计算机一样，变量数量的变化由这些电量的变化来表示。由于全电子模拟计算机没有移动部件，因此它们比机械和机电计算机更容易编程和维护。这些机器最终完全取代了机械和机电计算机，并在 20 世纪 70 年代成为模拟计算机的标准类型。

Purely mechanical analog computers eventually gave way to electromechanical and electronic computers. Variables in mechanical analog computers would be represented by quantities such as the left-right displacement of a shaft, the angle of a gear, or the running total number of rotations of a roller or drum: in other words, quantities of the problem to be solved were represented by physical quantities in the computing machine, and variation in those quantities was represented by physically moving those parts. Electromechanical and electronic computers used electrical quantities, such as voltage and resistance, to represent the quantity of a variable. And, like the mechanical analog computer, changes in the quantity of a variable were represented by a change in those electrical quantities. Because fully electronic analog computing machines had no moving parts, they were more easily programmed and maintained than their mechanical and electromechanical counterparts. These machines eventually replaced the mechanical and electromechanical entirely, becoming the standard type of analog computer by the 1970s.

这些模拟计算机解决的大多数问题本质上都是连续的，因此计算机用来表示变量的物理量本质上也是连续的（例如，角度、旋转和电压都自然被视为连续量）。但是，也存在重要的例外：一些研究的问题需要绝对值函数或阶跃函数等。在这些情况下，模拟计算机将使用不连续变量来表示这些不连续性（James 等人，1971 年；Maley，即将出版）。

The majority of the problems solved on these analog computing machines were continuous in nature, so the physical quantities that the machines used to represent variables were likewise continuous in nature (e.g., angles, rotations, and voltages are all naturally considered as continuous quantities). However, there are important exceptions: some of the problems studied would require, for example, the absolute value function, or step functions. In those cases, the analog computing machines would use discontinuous variables to represent those discontinuities (James et al., 1971; Maley, forthcoming).

尽管数字计算机现在已经完全取代了模拟计算机，但这一过程花了几十年的时间。最早的数字计算机比模拟计算机更大、更慢、更容易出错、更昂贵。但一旦它们处于同等地位，就必须向潜在买家和消费者区分这两种类型。用户。在那个历史时期，模拟计算机和数字计算机都是电子计算机。但是，模拟计算机（大多数）使用连续电压来表示变量，而数字计算机则完全使用离散电压。偶尔会出现模拟计算机使用不连续（即离散）电压的例外，但这种情况被忽略了。对于数字计算机来说，没有例外：它们总是使用离散电压。这是“模拟/数字”区别的一个重要来源：模拟机器是连续的（或者，至少，通常如此），而数字机器是离散的。

Although digital computers have now completely replaced analog computers, it took a few decades for them to do so. The earliest digital computers were larger, slower, more error-prone, and more expensive than their analog counterparts. But once they were on equal footing, it became important to distinguish the two types to potential buyers and users. At that point in history, both the analog and digital computing machines were electronic. However, the analog ones (mostly) used continuous voltages to represent their variables, while the digital ones used exclusively discrete voltages. The occasional exception where analog computers used discontinuous (i.e., discrete) voltages was overlooked. For digital computers, there were no exceptions: they always used discrete voltages. This was one important source of the “analog/digital” distinction: analog machines were continuous (or, at least, very often so), while digital machines were discrete.

关于这种划分的历史当然还有很多话要说，但这只是我们如何得出模拟/数字区别的公认观点的一部分。同样，根据这种观点，“模拟”与“连续”同义，因此模拟表示连续变化，模拟计算机只是使用连续表示的计算机。另一方面，“数字”与“离散”同义，因此数字计算机只是使用离散表示的计算机。

There is certainly more to say about the history of this division, but this is part of the story of how we arrived at the received view of the analog/digital distinction. Again, according to this view, “analog” is synonymous with “continuous,” and thus analog representations vary continuously, and analog computers are just computers that use continuous representations. On the other side, “digital” is synonymous with “discrete,” and thus digital computers are just computers that use discrete representations.

现在，即使电子模拟计算机确实总是使用连续电压（根据定义，传统观点就是如此），模拟计算机和数字计算机之间最重要的区别不仅仅是连续性与离散性的问题。更重要的是——我将在下一节中详细说明——电压如何表示。3简而言之^，模拟表示（以及模拟计算机）通过物理量的大小来表示数字的大小；数字表示（以及数字计算机）通过数字来表示数字的名称，然后数字由物理量的变化来表示。下面将对此进行更详细的讨论，但现在，我只想指出，这些计算机及其表示如何运作的这一关键区别被后来成为传统模拟/数字区别的东西所掩盖。传统观点掩盖了一个事实，即模拟和数字之间的差异比花生酱类型的差异更有趣，一种是块状的，另一种是光滑的。相反，有两种根本不同的表示方式，其中一种本质上利用了物理表示的物理性质，另一种则从物理表示中抽象出来。

Now, even if it were true that electronic analog computers always used continuous voltages (as the received view would have it, by definition), the most important difference between analog and digital computers is not simply a matter of continuity versus discreteness. More important—as I will spell out in detail in the next section—is how the voltages do the representing.³ In short, analog representations (and, in turn, analog computers) represent the magnitudes of numbers via the magnitudes of physical quantities; digital representations (and, in turn, digital computers) represent the names of numbers via digits, where the digits are then represented by variations in physical quantities. Much more will be said about this below, but for now, I simply note that this crucial difference in how these computing machines and their representations operate was glossed over by what became the received analog/digital distinction. The received view has obscured the fact that the difference between the analog and digital is more interesting than the difference between types of peanut butter, where one is chunky and the other is smooth. Rather, there are two fundamentally different ways of representing, one of which essentially takes advantage of the physical nature of physical representations, and another that abstracts away from them.

5.2 表现类型

5.2  Representational Types

根据我在此提供的框架，表征类型之间存在两个基本区别。首先，有模拟的，也有非模拟的。其次，有连续的，也有非连续的（即离散的）。4 根据这种观点（与传统观点相反），“模拟”并不^等同于“连续”，这一观点首先由 Lewis（1971 年）明确提出，随后由 Maley（2011 年）进行辩护和扩展（这一观点也隐含在 Copeland，1997 年的著作中）。最初拒绝这种同义词的理论原因已在现存（尽管是历史性的）模拟计算机的例子中得到证实（Maley，即将出版）。此外，“数字”并不等同于“离散”。相反，“数字”只是众多非模拟数字表示方式中的一种（尽管它的复杂性被它的熟悉性所掩盖）。在当代数字计算机中，正是这种特殊的表示类型（而不仅仅是离散性）对于理解其作为计算机的运作至关重要。

According to the framework I offer here, there are two fundamental distinctions to be made between representational types. First, there are those that are analog and those that are not. Second, there are representations that are continuous and those that are not (i.e., discrete).⁴ On this view (contra the received view), “analog” is not synonymous with “continuous,” a point first made explicit by Lewis (1971), then defended and extended by Maley (2011) (this point is also implicit in Copeland, 1997). What were originally theoretical reasons for rejecting this synonymy have been borne out in examples of extant (albeit historical) analog computers (Maley, forthcoming). Moreover, “digital” is not synonymous with “discrete.” Instead, “digital” turns out to be just one of many non-analog ways of representing numbers (although its complexity is obscured by its familiarity). In contemporary digital computers, it is this particular type of representation—and not mere discreteness—that is essential to understanding their operation qua computer.

现在，应该说有些哲学家已经为公认观点辩护。例如，Goodman (1968) 为公认观点的一个复杂版本辩护。Haugeland (1981) 也为公认观点辩护，将近似程序的工程考虑带入了讨论中。后来对公认观点或其变体的辩护可以在 (Papayannopoulos, 2020; Katz, 2016; Schonbein, 2014) 中找到。不过，现在我将把对这些观点的批评放在一边。

Now, it should be said that some philosophers have defended the received view. Goodman (1968), for instance, defends a sophisticated version of the received view. Haugeland (1981) also defends the received view, bringing engineering considerations of approximation procedures into the conversation. Later defenses of the received view, or variations thereof, can be found in (Papayannopoulos, 2020; Katz, 2016; Schonbein, 2014). For now, however, I will set critiques of these views aside.

在我们开始认真讨论之前，需要先了解一些术语。在下文中，我将把执行表示的物理事物简称为表示，将被表示的事物称为被表示物。这种区别与载体/内容区别的接近程度取决于如何理解“载体”和“内容”，这在文献中有所不同。有些人认为载体是物理的，而另一些人则不这么认为；有些人认为内容只是心理内容，而另一些人则不这么认为。为了避免可能的混淆，我将使用这些新术语。

A bit of terminology is needed before we begin in earnest. In what follows, I will refer to the physical thing doing the representing simply as the representation, and I will refer to what it is that is being represented as the representandum. How closely this distinction aligns with the vehicle/content distinction depends on how “vehicle” and “content” are understood, which varies in the literature. Some understand vehicles to be physical, and others do not; some understand content to be only mental content, and others do not. In order to avoid possible confusion, I will use these new terms.

上面做出的第一个区别——模拟表示和非模拟表示——反映了表示数字的两种不同方式。假设数学柏拉图主义是正确的——数字是真实的、抽象的对象。为了操纵这些抽象对象（就像我们习惯做的那样，无论是手工还是用计算机），我们必须以某种方式具体地表示它们。一种方法是通过它们的大小来表示数字：数字一由某个具体的物理量一表示；数字二由具体的物理量二表示；等等。因此，数字二可以用一根杆的长度来表示，其中杆长两米；或者可以用电路元件两端的电位来表示，其中电位为两伏。这是模拟的方式。

The first distinction made above—that between analog and non-analog representations—reflect two distinct ways of representing numbers. Suppose for a moment that mathematical Platonism is true—that numbers are real, abstract objects. In order to manipulate these abstract objects (as we are wont to do, both by hand and with computing machines), we must concretely represent them somehow or another. One way to do that is to represent numbers via their magnitudes: the number one is represented by some concrete, physical magnitude of one; the number two is represented by a concrete, physical magnitude of two; and so on. Thus, the number two could be represented by the length of a rod, where the rod is two meters long; or it could be represented by the electrical potential across a circuit element, where the potential is two volts. This is the way of the analog.

第二种（完全不同的）表示数字的方式是用我们所谓的名称来表示数字，名称有多种类型。有些名称是完全任意的。例如，“Drumthwacket”和“Graceland”是美国特定建筑物的任意名称；同样，“ e ”和“ π ”也是特定数字的任意名称。名称本身几乎没有提供有关这些名称所指对象的信息。或者，有些名称具有某种结构。如果我们将地址视为名称（有时确实如此），那么“123 Main St.”就是根据某些约定给出某个城市中特定房屋位置的名称。同样，“314”是根据某些约定给出某个进制数的值的名称。这是非模拟的方式。

The second—completely distinct—way to represent numbers is to represent them by what we might call their names, where names are of different types. Some names are completely arbitrary. For example, “Drumthwacket” and “Graceland” are arbitrary names for particular buildings in the United States; similarly, “e” and “π” are arbitrary names for particular numbers. Almost no information about the referent of those names is given by the names themselves. Alternatively, some names have a type of structure. If we consider addresses to be names (as they sometimes are), then “123 Main St.” is a name that gives the location of a particular house in some city or another, according to certain conventions. Similarly, “314” is a name that gives the value of a number in some base, according to certain conventions. This is the way of the non-analog.

简而言之，它们之间的重要区别在于，我们可以用数字的大小或名称来表示数字；这就是模拟表示和非模拟表示之间的区别（图 5.1中的左右划分）。这种差异反映在它们的物理表示方式上。就其物理实例而言，模拟表示和非模拟表示可以描述为一阶表示和二阶表示之间的差异（Maley，2021）。这种理解两者的方式有助于明确使用这些表示的计算类型之间的差异，下一节将对此进行讨论。

In short, the important difference between these is that we can represent a number either by its magnitude or by its name; this is the difference between analog and non-analog representation (the left-right division in figure 5.1). This difference is then reflected in how these are physically represented. With respect to their physical instantiations, analog and non-analog representation can be characterized as a difference between first-order and second-order representation (Maley, 2021). This way of understanding the two helps to make clear the differences between the types of computation that use those representations, discussed in the next section.

让我们更详细地了解每种表示方式，从模拟表示开始。我将论证模拟（或非模拟）是根本，然后我就可以说明为什么非模拟与数字不同。

Let us look at each of these ways of representing in more detail, starting with analog representation. I will argue that being analog (or not) is fundamental, which will then enable me to show why being non-analog is not the same thing as being digital.

5.2.1 模拟与非模拟表示

5.2.1  Analog vs Non-Analog Representations

因为我们在这里关注的是某种物理计算机（无论是自然的还是人造的），所以表征首先是某种物理对象。表征通过其物理的一个（或多个）属性；我认为，拼写“via”的不同方式正是区分表征类型的最根本方法。

Because we are concerned here with physical computers of some kind or another (either natural or artificial), representations are first and foremost physical objects of some kind or another. Representations do their representing via one (or more) of their physical properties; the different ways to spell out “via” is precisely what I argue is most fundamental about distinguishing representational types.

如上所述，表示表示数字的一种方式（模拟方式）是其属性之一的数量或量与被表示项共变。例如，可以使用一块粘土来表示数字，方法是使用粘土的温度（摄氏度）或粘土的质量（十克）。为了使示例具体化，假设我们要表示数字 50。5^因此，要通过粘土的质量表示 50，粘土的质量应为 50 十克。如果我们想将被表示项增加到 51，则需要将质量增加到 51 十克。因此，随着被表示项的变化，质量（执行表示的属性）也会随之变化。具体而言，被表示项的增加需要增加表示的质量；被表示项的减少需要减少表示的质量。因此，我们有一个数量（表示的属性之一）来表示被表示项的数量。⁶表示量的大小就是被表示数的大小；因此，这种表示方式的“类似”在于表示量与被表示数的大小之间的类似性。

As mentioned above, one way for a representation to represent a number—the analog way—is for the amount, or quantity, of one of its properties to covary with the representandum. For example, one can use a lump of clay to represent a number by using the temperature of the clay in degrees Celsius, or the mass of the clay in decagrams. To make the example concrete, let us suppose that we want to represent the number fifty.⁵ Thus, to represent fifty via the clay’s mass, the clay would have a mass of fifty decagrams. If we want to increase the representandum to fifty-one, we need to increase the mass to fifty-one decagrams. Thus, as the representandum changes, the mass—the property doing the representing—changes in turn. Specifically, an increase in the representandum requires an increase in the mass of the representation; a decrease in the representandum requires a decrease in the mass of the representation. Thus, we have a quantity—one of the properties of the representation—representing the quantity of the representandum.⁶ The magnitude of the representing quantity just is the magnitude of the number being represented; thus, what is “analog” about this type of representation is the analogy between the magnitude of the representation and the magnitude of the representandum.

这种特征的一个重要部分是，它可以适用于连续或离散的表示和被表示项。对于我们的粘土示例，假设我们只对表示整数感兴趣，并且假设我们只有一十克的粘土块可以添加到我们的表示中（也许我们表示的数字是一个人的年龄计数）。在这种情况下，我们仍然会有一个模拟表示，即使执行表示的属性和被表示项都以离散而不是连续的步骤变化。重要的不是表示或被表示项是否涉及连续性，而是执行表示的表示的属性是否与被表示项共变。在许多情况下，协方差将是线性的，但至少它必须是单调的：被表示项的量级的增加/减少以表示的相关属性的量级的增加/减少为特征^。7

An important part of this characterization is that it can hold for representations and representanda that are either continuous or discrete. For our clay example, suppose we are only interested in representing whole numbers, and suppose we only have single-decagram blobs of clay to add to our representation (perhaps the number we are representing is a count of a person’s age in years). In such a case, we would still have an analog representation, even though the property doing the representing and the representandum both vary in discrete, rather than continuous, steps. What matters is not whether continuity is involved in either the representation or representandum, but that the property of the representation that does the representing covaries with the representandum. In many cases, the covariance will be linear, but at a minimum it must be monotonic: an increase/decrease in the magnitude of the representandum is characterized by some increase/decrease in the magnitude of the relevant property of the representation.⁷

另一方面，非模拟表示则不是这样工作的。利用我们的粘土块，我们可以创建数字 50 的非模拟表示，方法是将粘土分成两个小块，将它们滚开，将其中一个做成数字“5”，将另一个做成数字“0”。按正确的顺序排列，这将表示数字 50。或者，我们可以将粘土块做成单个数字“L”，即罗马数字 50。8^现在，虽然表示的某些物理属性（或其模式）必须起到表示作用（否则物理表示怎么能表示任何东西？），但反映被表示项大小的不是起到表示作用的物理属性的大小。相反，表示的物理属性以某种方式抽象出来，因此重要的不是属性的大小，而是属性的任意值或模式随着被表示项的变化而变化。

Non-analog representation, on the other hand, does not work this way. Using our lump of clay, we could create a non-analog representation of the number fifty by separating the clay into two smaller lumps, rolling them out, and forming one into the numeral “5” and the other the numeral “0.” Put in the right order, this would represent the number fifty. Alternatively, we could form our lump of clay into the single numeral “L,” the Roman numeral for fifty.⁸ Now, while it must be true that some physical property (or pattern thereof) of the representation must do the representing (how else could a physical representation represent anything?), it is not the magnitude of the physical property doing the representing that reflects the magnitude of the representandum. Rather, the physical properties of the representation are abstracted in a way so that it is not the magnitude of the property that matters, but only variation among arbitrary values or patterns of the property as the representandum varies.

让我们再看一个例子，这次是在现代汽车仪表盘上常见的电子表示的背景下。图 5.2的左侧描绘了一个模拟表示，它将从代表 6 变为 7，右侧描绘了一个单一的数字、非模拟表示，它做同样的事情；我们可以想象它们中的每一个都代表汽车中剩余的燃油量（以加仑为单位）。请注意，模拟显示屏中的段可能表示显示屏由离散元素组成，或者它们可能只是用于帮助读取连续变化的（比如）液位（即，它可能是离散的或连续的）的标记。对于这里提出的模拟表示的特征，这并不重要。无论哪种方式，仪表（此处表示为灰色）的大小（在本例中为高度）决定了代表值。因此，高度的增加（或减少）对应于代表值的增加（或减少）。

Let us look at another example, this time in the context of electronic representations of the kind we often find on the dashboards of contemporary cars. The left side of figure 5.2 depicts an analog representation as it would change from representing six to seven, and the right side depicts a single numerical, non-analog representation doing the same thing; we can imagine each of these as representing, say, the remaining amount of fuel in a car (in gallons). Note that the segments in the analog display could indicate that the display is composed of discrete elements, or they may simply be marks meant to help read a continuously-varying level of (say) liquid (i.e., it may be discrete or continuous). On the characterization of analog representation advanced here, it does not matter. Either way, it is the magnitude (in this case, height) of the gauge (depicted here as gray) that determines the representandum. Thus, an increase (or decrease) in height corresponds to an increase (or decrease) in the representandum.

对于图 5.2右侧所示的数值表示，材料的大小或数量与表示本身无关。重要的是，某些部分以指定的模式被激活。随着被表示项从六增加到七，表示中的大小或数量不会增加。相反，被激活的元素只是从一种模式变为另一种模式。具体来说，段 D、E、F 和 G 被停用，段 B 被激活，段 A 和 C 保持不变。同样，在模拟表示的情况下，随着被表示项的增加（或减少），表示的量级会单调变化；在数字情况下，表示的不是任何物理属性本身的大小，而只是能够改变的物理量的模式（事实上，在这个例子中，当我们从六增加到七时，激活段的数量会减少，但如果我们从七增加到八，激活段的数量就会增加）。

As for the numerical representation depicted on the right side of figure 5.2, the magnitude or quantity of the material is irrelevant to the representation qua representation. What matters, instead, is that certain segments are activated in specified patterns. As the representandum increases from six to seven, there is no magnitude or quantity that increases in the representation. Instead, the elements that are activated simply change from one pattern to another. In particular, segments D, E, F, and G are deactivated, segment B is activated, and segments A and C are unchanged. Once again, in the case of the analog representation, there is a monotonic change in the magnitude that does the representing as the representandum increases (or decreases); in the digital case, it is not the magnitude of any physical property per se that does any representing, but only a pattern of a physical quantities that are capable of changing (in fact, the number of activated segments decreases in this example as we go from six to seven but would then increase if we went from seven to eight).

这些例子的教训具有普遍性。根据这种观点，任何模拟表示的定义特征是，表示的物理属性的量级或数量会随着被表示项的增加（或减少）而系统地（即单调地）增加（或减少）。在非模拟表示的情况下，表示的物理属性只需随着被表示项的增加或减少而以某种方式改变即可。当被表示项系统地变化时，表示的表示相关属性的量级不需要发生系统性变化。

The lesson from these examples generalizes. The defining feature of any analog representation on this view is that the magnitude or amount of the physical property that does the representing increases (or decreases) systematically (i.e., monotonically) with increases (or decreases) in the representandum. In the case of a non-analog representation, the physical property that does the representing only needs to change in some way or another as the representandum increases or decreases. There need be no systematic change in the magnitude of the representationally relevant property of the representation as the representandum systematically varies.

如图5.1所示，有些非模拟表示值得与数字区分开来。数字的特殊之处在于它们可以组合形成复数表示，而非数字则不然。非数字表示的例子包括“ π ”和“ e ”，它们代表数学常数 pi 和欧拉数。这些符号可以用与上述数字显示相同类型的分段显示来表示。但是，将它们称为数字表示是错误的：这些类型的符号不像数字那样连接在一起形成复数表示（即数字表示，将在下文中详细讨论）。

As illustrated in figure 5.1, there are some non-analog representations that are worth distinguishing from numerals. Numerals are special in that they combine to form complex representations, whereas non-numerals do not. Examples of non-numeral representations include “π” and “e,” which represent the mathematical constants pi and Euler’s number. These symbols could be represented with the same kind of segmented display as the numerical display above. However, it would be a mistake to call these digital: these kinds of symbols are not concatenated in the same ways that numerals are to form complex representations (i.e., numerical representations, which will be discussed in more detail below).

最后一点值得注意。虽然从概念上讲，连续的非模拟表示是不存在的。可以肯定的是，在某种意义上，连续表示存在非模拟表示；实数的数字展开就是一个例子。9^但是，没有任何表示本身在表示的属性上连续变化，但该属性不以模拟方式表示（因此图 5.1右下象限中的灰色框）。同样，许多表示具有连续变化的物理属性，但随后被用作非模拟表示。当代数字计算机中实现的二进制数字（用电压表示）就是一个例子。电压在不同值之间变化时连续变化（例如，从大约零伏的“低”值到大约五伏的“高”值），我们将该电压离散化（基本上将其四舍五入为零或五）并将其视为只有两个值：低值表示数字“0”，高值表示数字“1”。

One final point is worth noting. Although it may be conceptually possible, there are no continuous, non-analog representations. To be sure, there are, in a sense, non-analog representations of continuous representanda; the digital expansion of a real number is one example.⁹ However, there are no representations that themselves vary continuously in the property that does the representing, but where that property does not represent in an analog way (hence the gray box in the lower right quadrant of figure 5.1). Again, many representations have physical properties that vary continuously but are then used as non-analog representations. Binary numerals as they are implemented in contemporary digital computers, represented by voltage, are an example. The voltage varies continuously as it changes between different values (e.g., a “low” value of around zero volts to a “high” value of around five volts), we discretize that voltage (basically rounding it to zero or five) and treat it as though it only has two values: the low value represents the numeral “0,” and the high value represents the numeral “1.”

5.2.2 数值表示

5.2.2  Numerical Representations

某些非模拟表示最有趣的地方可能在于，它们可以以各种方式组合起来形成复杂的表示。其中包括我们熟悉的数字表示，这对于我们的目的来说最为重要，但还有其他一些表示也值得一提。根据 Chrisomalis (2020) 的说法，我将这些表示称为数值表示。数值表示有几种类型，都很有趣，但我将重点讨论数字表示（Chrisomalis 将其归类为一种累积位置数值系统）。从历史上看，这种表示方案甚至不是不同文化中最常见的。然而，鉴于该方案是当代数字计算的基础，它对于理解现存的计算系统具有特殊的哲学意义。

Perhaps the most interesting thing about certain non-analog representations is that they can be combined in various ways to form complex representations. Among these are the familiar digital representations, most important for our aims here, but there are others worth mentioning. Following Chrisomalis (2020), I will call these numerical representations. There are several families of numerical representations, all quite interesting, but I will focus the discussion here on digital representations (which Chrisomalis categorizes as a type of cumulative positional numerical system). Historically, this kind of representational scheme is not even the most common among different cultures. However, given that this scheme is the basis of contemporary digital computation, it is of particular philosophical importance for understanding extant computational systems.

在数字表示方案中，单个数字以不同的方式组合，有时还与辅助符号组合，以形成新数字的表示。例如，罗马数字方案认为每个数字都有特定的值，该值与每个数字本身的值相同。因此，“I”代表一；“V”代表五，依此类推。数字的书写方式是最大值在最左边，向右减小。整个数字串所代表的数字是各个数字所代表的数字之和。因此，“XVI”由代表十、五和一的数字组成；因此整个字符串代表十六。惯例中出现了轻微的复杂情况，即一个数字后面跟着一个较大的数字（在右边），这意味着要从较大的数字中减去较小的数字。因此，“XIV”代表十四。

In numerical representational schemes, individual numerals are combined in different ways, sometimes with auxiliary symbols, to form representations of new numbers. The Roman numeral scheme, for example, takes each numeral to have a certain value, which is the same value that each numeral has on its own. Thus, “I” represents one; “V” represents five, and so on. Numerals are written so that the largest-valued ones are left-most, and decrease to the right. The number represented by the entire string of numerals is the sum of the numbers represented by the individual numerals. Thus, “XVI” consists of the numerals representing ten, five, and one; the whole string thus represents sixteen. A slight complication arises with the convention in which a numeral followed by a larger numeral (on the right) was taken to mean that the smaller numeral was to be subtracted from the larger. Thus, “XIV” would represent fourteen.

另一个例子是希腊数字方案。在这个方案中，字母表中的前九个希腊字母代表一到九的值；接下来的九个字母代表十到九十的值；最后九个字母代表一百到九百的值。同样，所表示的数字是每个数字所表示的数字的总和。使用相同的从左到右、从大到小的惯例，“ ΦΛ B ”代表五百 ( Φ ) 三十 ( Λ ) 二 ( B )。有趣的是，在罗马和希腊等方案中，不需要数字来表示零；因此“DII”和“ Φ B ”分别代表五百零二。

Another example is the Greek numeral scheme. In this scheme, the first nine Greek letters of the alphabet represent the values one through nine; the next nine letters represent the values ten through ninety; the last nine represent the values one hundred through nine hundred. Again, the number represented is the sum of the numbers represented by each individual numeral. Using the same left-to-right, largest-to-smallest convention, “ΦΛB” represents five hundred (Φ) thirty (Λ) two (B). Interestingly, in schemes like the Roman and Greek, there is no need for a numeral to represent zero; thus “DII” and “ΦB” each represent five hundred and two.

还有其他方案使用乘法而不是加法，以及普通分数的除法方案；Chrisomalis (2020) 详细讨论了其中的许多方案。所有这些方案的共同点在于，它们的成员都是由单个数字构成的复合体，其中整个表示的代表项是单个数字的代表项的函数。因此，让我们转向最熟悉的数值方案，即数字方案，这也是我们对物理计算的几乎所有理解的基础。

There are still other schemes that use multiplication instead of addition, as well as the divisional scheme of common fractions; Chrisomalis (2020) discusses a number of these in fascinating detail. What unites all of these schemes is that their members are complexes, built out of individual numerals, where the representandum of the entire representation is a function of the representanda of the individual numerals. So let us turn to the most familiar numerical scheme, the digital scheme, which is also the basis upon which virtually all of our understanding of physical computation is built.

数字方案最早由刘易斯 (1971) 在哲学文献中提出，将数字表示为一串数字。每个数字对被代表数的贡献取决于其在字符串中的位置以及方案的进制。我们最熟悉的是使用十进制，它需要 0 到 9 的数字。最右边的数字（此处的“数字”是指特定位置的数字）为被代表数贡献了那么多个单位，下一个（左边的）数字贡献了那么多个十进制，依此类推。因此，“272”的被代表数等于：

The digital scheme, first articulated in the philosophical literature by Lewis (1971), represents a number as a string of numerals. The value that each numeral contributes to the representandum depends on its position in the string, as well as the base of the scheme. Most familiarly, we use base-10, which requires the numerals 0 through 9. The right-most digit (“digit” here is the term for a numeral in a particular place) contributes that many units to the representandum, the next (to the left) digit contributes that many tens, and so on. Thus, “272” has the representandum equal to:

更一般地，对于除 10 之外的基数^b，我们将上面的 10 替换为数字b，数字范围从 0 到 ( b -1)。有趣的是，数字方案可以通过使用小数位进一步扩展。小数位右边的数字可以表示基数的负幂，从而可以表示实数。

More generally,¹⁰ for a base b other than 10, we would replace the 10 above with the number b, and the numerals would range from 0 to (b− 1). Interestingly, the digital scheme can be further extended by the use of a decimal place. Digits to the right of the decimal place can then represent negative powers of the base, allowing for real numbers to be represented.

对于计算机来说，数字方案的一个有用事实是，只需两个数字即可表示任何可表示的数字。同样，由于我们关心的是物理表示和计算，这使得数字机器的实现（相对）简单：所需要的只是一个可以处于两种不同物理状态的物理基底。因此，二进制系统在当代数字计算机中几乎是通用的。

When it comes to computers, one useful fact about the digital scheme is that only two numerals are necessary to represent any representable number. Again, because we are concerned with physical representation and computation, this makes the implementation of a digital machine (relatively) simple: all one needs is a physical substrate that can be in two different physical states. The binary system is thus virtually universal in contemporary digital computers.

在结束本节之前，我们应该注意符号表示和数值表示之间的区别。符号表示根本不用于这里所理解的数值表示。我这样说的意思是，单个符号表示不会连接或以其他方式系统地与其他符号表示组合以形成新的表示。当然，这是一个偶然的事实：我们可以使用字符串“ π 0”来表示 ( π × 10 ¹ ) + (0 × 10 ⁰ ) = 31.4145926 ...但不存在像这样的现存表示方案。当然，符号的连接确实发生在数学环境中，但它们不是数值方案^。11

Before concluding this section, we should note the difference between symbolic representation and numerical representation. Symbolic representations are simply not used in numerical representations as understood here. By this I simply mean that individual symbolic representations are not concatenated or otherwise systematically combined with other symbolic representations to form new representations. This is, of course, a contingent fact: we could have used the string “π0” to represent (π × 10¹) + (0 × 10⁰) = 31.4145926… But there are no extant representational schemes like this. Concatenations of symbols do occur in mathematical contexts, of course, but they are not numerical schemes.¹¹

费尽心思创建这种表示类型的目的是为了展示模拟表示和数字表示之间的巨大差异。它们不仅仅是同一枚硬币的两面。相反，模拟表示和非模拟表示之间存在更根本的区别。数字是非模拟方面的一种特殊表示类型，而数字表示则是一种由数字（以及许多其他数字）构成的复杂表示类型。模拟表示的不同之处并不在于它们是连续的，而是在于它们以根本不同的方式表示数字。前面我提到，这种差异也可以描述为一阶表示与二阶表示；我们现在可以明白这意味着什么了。

The point of going to the trouble of creating this typology of representations is to show how very different analog and digital representation really are. They are not merely two sides of the same coin. Instead, there is a more fundamental division between representations that are analog and those that are not analog. Numerals are a particular type of representation on the non-analog side, and digital representation is then one particular complex representational type built of numerals (among many others). Analog representations are not different because they are continuous but because they represent numbers in a fundamentally different way. Earlier I mentioned that this difference can also be characterized as first-order versus second-order representation; we can now see what that means.

首先要使用一些物理属性作为表示（或表示的一部分）。如上所述，对于模拟表示，物理属性的量级表示被表示项的量级。在电子模拟计算机中，5 伏特表示数字 5。将其视为一阶表示。对于数字表示，物理属性的某个值表示一个符号（数字或其他符号），该符号表示被表示项的名称（或名称的一部分）。在电子数字计算机中，5 伏特¹²表示数字“1”。一般来说，数字“1”是某些数字表示的一部分，它表示个位数、两位数或四位数等等。相对于模拟表示，这是二阶表示。

Start with some physical property to be used as a representation (or part of a representation). As we saw above, with analog representation, the magnitude of the physical property represents the magnitude of the representandum. In an electronic analog computer, five volts represents the number five. Think of this as first-order representation. For digital representation, some value of the physical property represents a symbol (a numeral or some other symbol), and that symbol represents the name of the representandum (or part of the name). In an electronic digital computer, five volts¹² represents the numeral “1.” In general, the numeral “1” is then a part of some digital representation, where it represents either the ones digit, or the twos digit, or the fours digit, and so on. Relative to analog representation, then, this is second-order representation.

尽管存在这些差异（现有观点无法仅使用“离散与连续”的资源来捕捉这些差异），但模拟和数字表示的相似之处在于，它们都是不同类型计算的基础。理解模拟和数字计算需要以本节所示的方式理解模拟和数字表示：这不仅仅是连续和离散表示之间的差异。如果我们接受计算需要表示的假设，那么我们只关注数字表示就对自己不利了，因为数字表示在表示领域中只占了相对较小的空间。模拟空间被忽略了，主要是因为人们误解了它的有趣特征。在下一节中，我们将开始纠正这种情况。

Despite these differences—differences that the received view cannot capture, using only the resources of “discrete versus continuous”—analog and digital representation are alike in that both are the basis of different types of computation. Understanding analog and digital computation requires understanding analog and digital representation in just the way shown in this section: it is not merely the difference between continuous and discrete representations. If we accept the assumption that computation requires representation, we have been doing ourselves a disservice by looking only at digital representations, which occupies a relatively small space of the representational landscape. The analog space has been ignored, largely because its interesting features have been misunderstood. In the next section, we will begin to remedy this situation.

5.3 计算类型

5.3  Computational Types

就像模拟和数字表示的情况一样，认为模拟和数字计算是根据它们使用的表示（或机制、变量或其他任何东西）是连续的还是离散的来区分的，这是错误的。相反，计算机器是模拟的还是数字的，只是它使用的是模拟还是数字表示，具体含义见第 5.2 节。这需要一些解析，这也是本节的目的。我将简要介绍一些数字和模拟计算的例子，并提到一些使它们独一无二的特点。

Just as it was for the case of analog and digital representation, it is a mistake to think that analog and digital computation are separated by whether the representations (or mechanisms, variables, or anything else) they use are continuous or discrete. Instead, what makes a computational machine analog or digital is just whether it uses analog or digital representations, in the specific sense outlined in section 5.2. This requires some unpacking, which is the purpose of this section. I will briefly examine some instances of digital and analog computation and mention some of the features that make each unique.

5.3.1 数字计算

5.3.1  Digital Computation

几乎所有当代计算机都是数字计算机，基于冯·诺依曼（1982 年）最初提出的设计。与流行的说法相反，这些计算机之所以特别数字化，是因为它们以多种方式利用了上述意义上的数字。关于计算背后的一般思想已经在其他地方讨论过了，所以这里让我们关注数字计算的具体数字化部分。所有数据最终都以数字形式存储，包括用于控制机器精确操作的程序。例如，数字计算机中的加法和乘法是通过我们从小学就熟悉的​​算法进行的：这些运算是逐位操作的结果。例如，数字的加法首先从添加最低有效位开始，然后添加次高有效位（如果需要，加上进位位），然后添加次高有效位，依此类推。这种算法只有使用我们都熟悉的数字系统才有可能——唯一的区别是，数字计算机使用的是二进制系统，而不是十进制。

Virtually all contemporary computing machines are digital computers, based on a design originally proposed by von Neumann (1982). Contrary to popular accounts, what makes these computers specifically digital is the many ways that they take advantage of numbers represented digitally, in the sense presented above. Much has been written about the general ideas behind computation elsewhere, so here let us focus on what is specifically digital about digital computation. All data is ultimately stored as numbers represented digitally, including the very programs that are used to control the precise actions of the machine. So, for example, addition and multiplication in digital computers happens via algorithms familiar to us from elementary school: these operations are the result of digit-by-digit manipulations. The addition of numbers, for example, starts with adding the least-significant digits first, then adding the next-most-significant digits (plus a carry digit, if necessary), then the next-most, and so on. This kind of algorithm is only possible using the digital system with which we are all familiar—the only difference is, in a digital computer, a base-2 system is used, rather than base-10.

同样，数字表示可以系统地组织数据。举个简单的例子，假设我们有 32 个可能的位置来存储数据。在二进制中，我们可以将这 32 个位置写为地址，编号为 00000、00001、00010、…… 11111。出于各种原因，将这些位置分散到几个不同的组中效率更高。因此，假设我们有四个不同的“芯片”，每个芯片有八个单独的位置。

Similarly, digital representations enable a systematic organization of the data. As a toy example, suppose we have thirty-two possible locations for storing data. In base-2, we can write those thirty-two as addresses numbered 00000, 00001, 00010, …11111. For a variety of reasons, it is more efficient to spread these locations among several different groups. So, suppose we have four different “chips,” each of which has eight individual locations.

当我们需要访问特定地址时，我们可以使用两个最高有效数字来定位特定芯片，然后使用三个最低有效数字来定位该芯片内的特定地址。数字计算机中的电路设计为数据地址的解析方式（尽管在实践中，解析的规模要大得多）。同样，这是一种利用数字表示中数字的特定组织方式的方法。

When we need to access a particular address, we can use the two most-significant digits to locate a particular chip, then the three least-significant digits to locate a particular address within that chip. The circuitry in a digital computer is designed so that this is how the addresses of data are parsed (although, in practice, at a much larger scale). Again, this is a way of taking advantage of the particular organization of numerals in a digital representation of numbers.

实际的数字计算机包含大量数据，这些数据通过大量操作以极高的速度进行处理。它们通过允许程序员以多种复杂的方式将少量操作串联起来来实现其功能，通常对于可能数据的不同组合使用不同的路径。然而，从根本上讲，它们的操作依赖于其原理与此处的玩具示例并无不同之处的组件；实际机器的唯一区别在于它们的规模。

Actual digital computers contain huge amounts of data that are processed using a large number of operations at very high speeds. They accomplish what they do by allowing programmers to string together a small number of operations in many complicated ways, usually with different paths for different combinations of possible data. At bottom, however, their operation depends on components whose principles are not different in kind to the toy example here; the only difference in actual machines is their scale.

数字计算有优点也有缺点。一个非常简单的例子是某些操作速度很快。在十进制中，增加数字的数量级就像在其数字表示中添加零一样简单。对于二进制，添加零将使数字乘以二。此外，我们可以通过简单地添加更多数字来提高任意表示数字的精度。但是，有些操作很慢。确定列表中的最大元素需要检查该列表中的每个元素。数字计算机必须以与人们必须在一叠索引卡片中找到最大的三位数的方式执行此操作，其中每张卡片上都写有一个数字。为了找到最大的数字，我们必须检查第一个数字并将其与第二个数字进行比较，保留两者中较大的一个。然后，我们必须将该数字与第三个数字进行比较，保留两者中最大的一个。对堆栈中的每一张卡片重复此过程后，我们将找到最大的元素。

There are advantages and disadvantages to digital computation. One very simple example is that certain operations are fast. In base-10, increasing the order of magnitude of a number is as simple as adding a zero to its digital representation. For base-2, adding a zero multiplies the number by two. Additionally, we can increase the precision with which we represent a number arbitrarily by simply adding more digits. However, some operations are slow. Determining the largest element in a list requires examining every element in that list. A digital computer must do this in much the same way a person would have to find the largest three-digit number in a stack of index cards, where each card has a single number written on it. In order to find the largest, we would have to examine the first one and compare it to the second, keeping the larger of the two. Then, we would have to compare that one to the third one, keeping the largest of those two. After repeating this process for every card in the stack, we will have found the largest element.

人们可能会想，我们用数字计算机做的所有事情似乎都与数字的操作或存储无关。我正在用文本编辑器写这句话，后台还有其他几个程序在运行，检查电子邮件、显示即将到来的约会的通知以及播放音乐。这些都与处理数字无关。

One might wonder about all of the things we do with digital computers that seem to have nothing to do with the manipulation or storage of numbers. I am writing this sentence with a text editor, and in the background, several other programs are running, checking email, displaying notifications about upcoming appointments, and playing music. Very little of this is about processing numbers.

答案当然是，这些东西最终都是用二进制数字在机器中表示的。不久前，CD 是一种广泛使用的音乐录音格式。但 CD 不仅仅可以保存音乐：它们可以保存任何类型的数据。这是因为，无论数据类型如何，它最终都表示为数字，以二进制表示，其中各个数字在物理上表示为 CD 表面上的凹坑（1）和凹岸（0）。然后，这些数字可以依次表示声音元素、字母、像素或任何其他特定程序已编写以将其解释为的东西。但存储、读取和操作所有这些东西的机制只有在数据以数字表示的背景下才能理解。

The answer, of course, is that these things are all ultimately represented in the machine by binary numbers. Not long ago, CDs were a widespread format for musical recordings. But CDs need not hold only music: they can hold data of any type. What makes this true is that, whatever the type of data, it is ultimately represented as numbers, which are represented in binary, and where the individual numerals are physically represented as pits (1s) and lands (0s) on the surface of the CD. Those numbers can then, in turn, represent elements of sounds, or letters, or pixels, or anything else that a particular program has been written to interpret them as. But the mechanisms that store, read, and manipulate all of those things can only be understood in the context of that data being represented as numbers represented digitally.

5.3.2 模拟计算

5.3.2  Analog Computation

数字计算使用数字表示，而模拟计算使用模拟表示。不幸的是，我们今天对模拟计算不太熟悉，但简单的例子可以很好地说明这些原理。

Whereas digital computation uses digital representations, analog computation uses analog representations. Unfortunately, we are today much less familiar with analog computation, but simple examples can illustrate the principles quite well.

首先，要了解模拟加法与数字加法有何不同，请再次考虑我们如何将一对两位数相加，就像我们在小学学到的那样。我们将它们排成一行，将个位数字相加并记下结果，然后添加十位数字，必要时包括进位数字。同样的加法运算可以以模拟方式完成使用两把长尺。假设我们要将三十四加到四十二。我们在第一把尺子上找到四十二，在第二把尺子上找到三十四。然后，我们滑动第二把尺子，使其零点与四十二对齐。最后，我们看看第一把尺子上的数字与第二把尺子上的三十四对应。从文字上看，这听起来很复杂，但我们要做的是将两个表示为量级的数字相加。因为表示的量级是长度，所以我们只需将这些长度相加（即连接起来）。

First, to understand how very different analog addition is from digital addition, consider again how we add a pair of, say, two-digit numbers as learned in elementary school. We line them up, add the ones digits and write down the result, then add the tens digits, including a carry digit if necessary. The same addition operation could be done in an analog fashion using two long rulers. Suppose we want to add thirty-four to forty-two. We locate forty-two on the first ruler, and thirty-four on the second. Then, we slide the second ruler so that its zero point lines up with forty-two. Finally, we see what number on the first ruler corresponds to thirty-four on the second ruler. Verbally this sounds complicated, but what we are doing is adding two numbers represented as magnitudes. Because the magnitudes that are doing the representing are lengths, we simply add (i.e., concatenate) those lengths.

20 世纪早期的机械模拟计算机使用的原理与此类似：数字用长度或其他量值（如角度、位移或不同物理部件的旋转次数）来表示。一个更复杂的例子是磁盘积分器，如图 5.6所示（改编自 Maley，即将出版）。

Mechanical analog computers of the early twentieth century used principles just like this: numbers were represented by lengths or other magnitudes such as angles, displacements, or the number of rotations of different physical components. A more complicated example is the disk integrator, shown in figure 5.6 (adapted from Maley, forthcoming).

假设我们要对某个函数进行积分：我们输入函数的值，然后输出该函数在该时间点之前的定积分。该积分器通过一个不断旋转的转盘（圆盘 A）工作，该转盘驱动一个垂直圆盘（B），该圆盘可以靠近或远离 A 的中心。当 B 靠近 A 的边缘时，它会相对快速地旋转；当它位于 A 的中心时，它根本不会旋转。

Suppose we want to integrate some function: we input the value of the function, and get as output the definite integral of that function up to that point in time. This integrator works via a constantly-rotating turntable (disk A), which drives a perpendicular disk (B), which can move closer or farther from the center of A. When B is near the edge of A, it will rotate relatively quickly; when it is at the center of A, it will not rotate at all.

该积分器的输入（要积分的函数值）是圆盘 B 相对于圆盘 A 中心的位移。输出是 B 旋转的总次数。如果要积分的函数从零开始然后增加，B 将从 A 的中心开始向外移动（同时 A 以恒定速度旋转）。如果函数减小，B 会向 A 的中心移动（如果函数变为负值，B 会移动到 A 的另一侧）。B 速度的变化使得 B 旋转的累计次数是函数的定积分。

The input to this integrator—the value of the function to be integrated—is the displacement of disk B relative to the center of disk A. The output is the total number of rotations of B. If the function to be integrated started at zero and then increased, B would begin in the center of A and move outward (all while A is rotating at a constant speed). If the function decreased, B would move back toward the center of A (and would move to the other side of A if the function became negative). This variation in B’s speed is such that the running total of B’s rotations is the definite integral of the function.

与数字计算机非常相似，这些组件可以以多种方式组合，以比人类所能执行的任何操作都要快得多的速度进行更复杂的计算。图 5.7显示了这种机械模拟计算机的一个例子，其中许多这样的组件连接在一起，以及可以以图形方式绘制输出或输入的绘图表。这些机械机器最终让位于电子模拟计算机，电子模拟计算机使用电压或电阻等电量来表示数字；Maley（即将出版）包含几个例子。

Much like digital computers, these components could be combined in a large number of ways to produce much more complicated computations at speeds greater than anything a human could perform. One example of such a mechanical analog computer is shown in figure 5.7, where many such components are connected together, as well as drawing tables upon which output or input could be plotted graphically. These mechanical machines eventually gave way to electronic analog computers that used electrical magnitudes, such as voltage or resistance, to represent numbers; Maley (forthcoming) contains several examples.

就像数字计算机一样，模拟计算机也有优点和缺点。有些操作很快。例如，假设（再次）我们要对一组数字进行排序。这一次，它们不是写在索引卡上（即以数字表示），而是用干意大利面条的长度来表示（示例取自 Dewdney，1984 年）。为了找到最大的元素，只需将所有面条捆在一起，将一端压在硬表面上，然后检查哪根面条伸得最高。这个过程本质上是一个步骤，而数字搜索任务需要与要搜索的元素一样多的步骤。其他操作较慢。虽然在数字表示中增加数量级相当于在表示中添加一位数字，但在模拟情况下，这种增加可能要复杂得多。在 1 厘米的长度上增加 9 厘米需要一些工作，再增加 90 厘米则需要更多的工作；在数字情况下，我们只需添加一位数字，然后再添加一位数字。

Just like their digital counterparts, analog computers have advantages and disadvantages. Some operations are fast. For example, suppose (once again) that we want to sort a set of numbers. This time, instead of being written down on index cards (i.e., represented digitally), they are represented by lengths of dry spaghetti noodles (an example taken from Dewdney, 1984). In order to find the largest element, one simply needs to hold the bundle of all noodles together, tamp one end down on a hard surface, and then examine which noodle sticks up the highest. This process is essentially a single step, whereas the digital searching task took as many steps as there were elements to search. Other operations are slower. Whereas increasing the order of magnitude in a digital representation amounted to adding a digit to the representation, such an increase in the analog case can be much more complicated. Adding nine centimeters to a length of one centimeter takes some work, and adding ninety more takes even more work; in the digital case, we simply add a digit, then add another one.

模拟计算的最大缺点可能是精度问题。因为模拟表示只是用物理量来表示量，所以我们对表示的了解仅限于我们可以观察到的物理量。例如，如果没有专门的设备，很难测量千分之一厘米的长度；因此，使用长度的表示可能仅限于少数有效数字。然而，在数字情况下，我们只需在表示中添加更多数字，就可以无限提高精度。

Perhaps the most significant disadvantage of analog computation is the issue of precision. Because analog representation just is the representation of magnitudes by physical magnitudes, what we can know about a representation is limited to what we can observe about the physical magnitudes in question. It is difficult to measure lengths in thousandths of a centimeter without specialized equipment, for example; thus, a representation using length might be limited to a small number of significant digits. In the digital case, however, we can increase precision indefinitely by just adding more digits to the representations.

除了博物馆和私人收藏外，模拟计算机已不复存在；这才是应该的。作为实用的计算机器，当代数字计算机几乎在各个方面都胜过模拟计算机。然而，模拟计算确实在我们对一般计算的理解中占有一席之地，而不仅仅是在数字的特殊情况下。此外，在许多情况下，神经系统可能使用模拟表示和计算。一个简单的例子是通过神经放电率来表示刺激：近一个世纪以来，人们已经知道，随着音调变大或肌肉拉伸越来越大，一些神经元会更频繁地放电（Adrian，1926）。因此，随着相关刺激的增加，放电率单调增加，这是一种模拟表示。将其与处理该表示的正确机制相结合，我们就有了一个明显的模拟计算案例（Maley，2018）。

Analog computers no longer exist in any quantity outside museums and private collections; this is how it should be. As practical computing machines, contemporary digital computers outperform analog computers in almost every way. However, analog computation does have a place in our understanding of what computation is in general, not just in the special case of the digital. Moreover, it may well be that, in many cases, neural systems use analog representation and computation. One simple example is the representation of stimuli via neural firing rates: it has been known for almost a century that some neurons will fire more frequently as a tone gets louder or a muscle is increasingly stretched (Adrian, 1926). The firing rate thus monotonically increases as the relevant stimulus increases, which is an analog representation. Combine this with the right mechanisms in place to process that representation, and we have a clear case of analog computation (Maley, 2018).

5.4 模拟对反计算主义的诉求

5.4  The Analog Appeal to Anti-Computationalism

正确理解模拟计算和表征，我们就能对心智（无论是自然的还是人工的）的计算性有了新的认识。这种方法可能会使心智/大脑首先是计算性的论点更容易被认知科学中的某些反计算/反表征研究项目接受（Isaac，2018 年提出了类似的观点）。特别是，激进具身认知（Chemero，2009 年；Van Gelder，1995 年）和激进行为认知（Hutto，2006 年）的某些支持者认为，认知不是计算，表征在理解认知中没有地位。然而，他们所考虑的都是数字（或至少是非模拟）计算，其中表征独立于其物理基础，系统可以用纯动态术语来理解，而不需要任何表征语言。

Properly understood, analog computation and representation give us a new understanding of what it could mean for minds—whether natural or artificial—to be computational. This approach might make the thesis that the mind/brain is computational in the first place more palatable to certain anti-computation/anti-representation research programs in cognitive science (a similar point has been made by Isaac, 2018). In particular, certain proponents of radical embodied cognition (Chemero, 2009; Van Gelder, 1995) and radical enactive cognition (Hutto, 2006) have argued that cognition is not computation, and that representation has no place in understanding cognition. However, what they all have in mind is digital—or at least non-analog—computation, where representations are independent from their physical bases, and systems can be understood in purely dynamical terms, without the necessity of any representational language.

同样，要说的比篇幅要多，但也许简短而错误（但希望是有趣的错误）比什么都不说要好。考虑一下图 5.8中描绘的瓦特调速器，这是范·盖尔德（1995 年）最初提出的例子，据称它是一个具有复杂行为的系统，可以在不设定表示的情况下完美描述。瓦特调速器是一种用于调节（或控制）蒸汽机速度的装置。蒸汽流入发动机，然后驱动飞轮，飞轮为其他机械提供动力。进入发动机的蒸汽越多，飞轮转动得越快；调速器通过调节流入发动机的蒸汽量，将速度保持在一个狭窄的范围内。飞轮连接到图 5.8左下方的水平滑轮，使金属球旋转。这些球通过铰链连接到臂，因此，当它们旋转时，它们会向外移动（通过所谓的离心力）。这些铰链连接到较小的臂，较小的臂连接到一个大垫圈，这样，当球旋转得更快时，它们会使垫圈向下移动，从而拉下连接到蒸汽管内阀门的杠杆，从而减少流过管道的蒸汽量。简而言之，它使用发动机速度来驱动一个机制，当速度过高时，该机制会降低速度（通过关闭蒸汽阀），但当速度过低时会增加蒸汽流量。它本质上是一种反馈机制，使蒸汽流量保持在一定范围内，从而使发动机保持在一定的转速。

Again, there is more to say than there is space to say it, but perhaps it is better to be brief and wrong (but interestingly wrong, one hopes) than to say nothing at all. Consider the Watt governor depicted in figure 5.8, an example originally introduced by Van Gelder (1995), purportedly as a system with complex behavior that can be perfectly described without positing representations. The Watt governor is a device meant to regulate (or govern) the speed of a steam engine. Steam flows into an engine, which drives a flywheel, which powers some other machinery. The more steam that goes into the engine, the faster the flywheel turns; the governor keeps this speed within a narrow range by adjusting the amount of steam flowing into the engine. The flywheel connects to the horizontal pulley on the bottom left of figure 5.8, causing the metal spheres to rotate. Those spheres are connected by hinges to arms so that, as they rotate, they move outward (by so-called centrifugal force). Those hinges are connected to smaller arms, connected to a large washer such that, as the spheres rotate faster, they cause the washer to move down, which in turn pulls down a lever connected to a valve within the steam pipe, which then reduces the amount of steam flowing through the pipe. In short, it uses the engine speed to drive a mechanism that then reduces that speed (by closing the steam valve) when the speed gets too high, but increases the steam flow when the speed gets too low. It is essentially a feedback mechanism that keeps the steam flow within a certain range, thus keeping the engine at a certain speed.

众所周知，范·盖尔德认为，臂角（理解这种装置的关键变量）可以完全用微分方程来表征。如果阀门是用一台小型数字计算机控制的，并连接传感器和伺服装置，那么我们当然必须讨论表示和算法才能理解这种装置。但事实并非如此，一个方程（具有正确的参数）完全涵盖了我们需要了解的有关该系统的所有信息。尽管人们可能倾向于认为臂角代表发动机转速，但范·盖尔德声称这是一个错误：方程就是我们所需要的一切，讨论表示和计算是多余的，甚至可能误导人。因此，该论点认为，我们的思维/大脑很可能是一个像瓦特调速器这样的系统，人们不需要讨论任何表示或计算，因此认知科学核心的计算假设可能是错误的。

Famously, van Gelder argued that the arm angle (the crucial variable in understanding this device) can be completely characterized by a differential equation. Had the valve been controlled with a small digital computer, hooked up to sensors and servos, we would certainly have to talk of representations and algorithms in order to understand the device. But given that it is not, a single equation (with the right parameters) completely captures all we need to know about this system. Although one might be tempted to think of the arm angle as representing the engine speed, van Gelder claims this would be a mistake: the equation is all one needs, and talk of representation and computation is superfluous and even potentially misleading. Thus—the argument goes—our mind/brain may well be a system like the Watt governor, where one does not need any talk of representation or computation, and so the computational assumption at the heart of cognitive science might be wrong.

有趣的是，瓦特调速器几乎是上述模拟计算机的完美例子（van Gelder 没有考虑到这种可能性），实际上与前面提到的二战期间和二战前后军舰上使用的机械模拟计算机非常相似（Mindell，2002）。这些机器将船的速度和航向作为输入，以轴的旋转和机械臂的角度表示，加上目标船的速度和航向信息，也以机械量表示。输出——击中目标所需的火炮位置——是输入的数学函数，由各种机械设备计算，输出为与相关火炮直接相关的物理位移。完全不清楚为什么这些机械模拟设备被视为代表输入和计算功能（根据创造它​​们的人的说法），而瓦特调速器却不是。

Interestingly, the Watt governor is almost a perfect example of an analog computer on the account I have outlined above (a possibility not considered by van Gelder), and actually quite similar to the mechanical analog computers used in military ships in and around World War II, mentioned earlier (Mindell, 2002). These machines took as input the ship’s speed and heading, represented as the rotation of shafts and angles of mechanical arms, plus information about the target ship’s speed and heading, also represented as mechanical quantities. The output—the gun position required to hit the target—was a mathematical function of the inputs, computed by various mechanical devices, and output as a physical displacement that connected directly to the relevant gun. It is not at all clear why these mechanical analog devices count as representing inputs and computing functions (by the lights of the very people who created them), whereas the Watt governor does not.

当然，瓦特调速器的臂角是否代表发动机转速是有争议的，而简单地断言它确实如此就会引出对范盖尔德的质疑。需要一个积极的论据，并且该论据不会最终将所有机械设备都描述为计算机。不幸的是，这样的论据会让我们走得太远。然而，可以证明的是，如果瓦特调速器在稍微不同的环境中使用，那么它将是一台模拟计算机，我们会将表示归因于它，以便分析和解释它的作用，即使按照范盖尔德的观点。

Of course, whether the arm angle of the Watt governor represents the engine speed is contentious, and simply asserting that it does would beg the question against van Gelder. A positive argument is required, and one that does not end up characterizing all mechanical devices as computational. Unfortunately, such an argument would take us too far afield. However, what can be shown is that, if the Watt governor were used in a slightly different context, then it would be an analog computer, and we would attribute to it representations in order to analyze and explain what it does, even by van Gelder’s lights.

如果我们使用发动机的转速作为输入，使用蒸汽阀的角度作为输出，那么瓦特调速器就是一台模拟计算机，其中计算的函数只是两者之间的数学关系。在这种情况下，它也与图 5.6中所示的机械积分器非常相似。现在，按照范盖尔德自己的看法，描述输入和输出之间关系的函数是一个数学函数，因此该设备将精确计算该函数。剩下的就是让输入和输出成为真实的表示。正如我们所见，各种物理量都可以作为模拟计算机的输入和输出：除了我们已经看到的电气和机械示例之外，经济模拟的 MONIAC 模拟计算机使用液位和液流作为变量（Isaac，2018）。因此，发动机转速和气门角度与任何东西一样好。

The Watt governor would be an analog computer if we were to use the speed of the engine as the input and the angle of the steam valve as the output, where the function computed is just the mathematical relation between the two. In that context, it would also be quite similar to the mechanical integrator illustrated in figure 5.6. Now, by van Gelder’s own lights, the function that describes the relationship between the input and output is a mathematical one, and thus this device would compute precisely that function. All that remains is for the inputs and outputs to be genuine representations. As we have seen, all kinds of physical quantities can serve as the inputs and outputs to analog computers: besides the electrical and mechanical examples we have already seen, the economy-simulating MONIAC analog computer used fluid levels and fluid flow as variables (Isaac, 2018). Thus, engine speed and valve angle are as good as anything.

现在，有人可能会反对我们可以随意取一个机械部件并将其用作模拟计算机的一部分的想法。然而，这并不奇怪。尽管模拟计算机并不像非模拟计算机那样独立于介质，但仍有各种各样的方法可以物理地创建模拟计算机。主要区别在于，由于给定机制所能表示的内容的物理限制，它只能实现一种特定的数学函数。例如，瓦特调节器的输出是其输入的特定三角函数，范围很窄。如果我们需要这样的功能，瓦特调节器将在我们的机械模拟计算机中完美运行，我们将通过展示臂角（和其他组件）如何表示方程中指定的量来解释将输入转换为输出的机制。但我们不能使用瓦特调节器来实现我们想要的任何功能。

Now, one might object to the idea that we can just take an arbitrary mechanical component and use it as part of an analog computer. However, this should not be all that surprising. Even though analog computers are not medium-independent in the way non-analog computers are, there is still a wide variety of ways to physically create an analog computer. The primary difference is that a given mechanism might only be able to implement one particular mathematical function because of physical constraints on what it can represent. The Watt governor, for example, has an output that is a specific trigonomic function of its input with a narrow range. If we need such a function, the Watt governor will work perfectly in our mechanical analog computer, and we will explain the mechanism that transforms the input to the output by showing how the arm angles (and other components) represent the quantities specified in the equation. But we cannot use the Watt governor to implement just any function we want.

激进的具身认知科学家和行动认知科学家是否会接受这些想法最终取决于他们自己；人们可能会认为，他们将会把这里提出的“模拟表征”概念看作一个相当单薄的表征概念，这反过来又使得模拟计算本身变得相当单薄。然而，如果历史先例真的重要的话，那么在数字计算机之前出现的模拟计算机，与任何其他计算形​​式一样，都是合法的计算形式。那些对计算怀有敌意但对动态系统友好的人（例如）可能不得不重新评估他们认为的敌人是谁。

Whether radical embodied and enactive cognitive scientists will be amenable to these ideas will ultimately be up to them; one might anticipate that they will take the notion of “analog representation” developed here to be a rather thin notion of representation, which in turn makes analog computation itself rather thin. However, if historical precedent matters at all, analog computers, which preceded digital computers, are just as legitimate a form of computation as anything. Those hostile to computation but friendly to dynamical systems (for example) may have to reevaluate who they think their enemies are.

当然，认知科学将思维/大脑视为计算机是否正确还有待观察。计算哲学家在帮助阐明计算可能是什么方面已经取得了很大进展，使这一论点上升到可检验假设的水平，而不是一个工作假设或一个纯粹的隐喻框架。正确理解模拟计算是这一努力的一部分，它为研究人员提供了更多选择，以原则性、系统性的方式理解思维/大脑如何成为计算机。

Of course, it remains to be seen whether cognitive science is correct in taking the mind/ brain to be a computer in the first place. Philosophers of computation have gone a long way in helping to clarify what computation might be, allowing for this thesis to rise to the level of a testable hypothesis, rather than a working assumption, or a mere metaphorical framework. Properly understanding analog computation is a part of this effort, allowing researchers more options for understanding how the mind/brain might be a computer in a principled, systematic way.

第 5 章注释

Notes for Chapter 5

1.   1.这并不是说没有写任何东西：一些很好的例子包括 Mahoney 的作品，以及（Campbell-Kelly 等人，2013 年）和（Haigh 和 Ceruzzi，2021 年）。

2.   1.  Which is not to say that nothing has been written: a few excellent examples include Mahoney’s work, as well as (Campbell-Kelly et al., 2013) and (Haigh and Ceruzzi, 2021).

3.   2.众所周知，“计算机”一词指的是进行计算的人；在本节中，我将使用“人类计算机”或“计算机器”来避免歧义。

4.   2.  It is well known that the term “computer” referred to a human performing computations; in this section, I will refer to either “human computers” or “computing machines” to avoid ambiguity.

5.   3 . 或者如果不是电压，无论实施媒介是什么。

6.   3.  Or if not voltages, whatever the implementational media happens to be.

7.   4.所谓“连续”表示，我指的是可以连续变化的表示：沿着连续体。单个表示本身不能是连续的；但用其他方式表述会比较麻烦，所以我将使用这种惯例。

8.   4.  By a “continuous” representation, I mean a representation that can vary continuously: along a continuum. An individual representation cannot, by itself, be continuous; but phrasing things otherwise is rather cumbersome, so I will use this convention.

9.   5.在这里，我将使用稍微麻烦一些的英语数字拼写，只是为了避免在文中使用一种表示方案来强调一种完全不同的方案。

10.   5.  I will use the somewhat cumbersome English spelling of numbers here, simply to avoid using one representational scheme in the text to make a point about a completely different scheme.

11.   6.请注意，我并没有具体说明表示的相关属性实际上会如何变化；这与此无关。目前，我只是阐明了要使某事物成为模拟表示，需要什么条件。

12.   6.  Note that I am not specifying how the relevant properties of the representation would actually change; that is not relevant here. For the moment, I am just elucidating what is required for something to be an analog representation at all.

13.   7.这在（Maley，即将出版）中得到了更精确的描述。

14.   7.  This is characterized more precisely in (Maley, forthcoming).

15.   8.第一个例子也是数字化表示的情况，尽管第二个例子不是：下一节将详细介绍这种区别。

16.   8.  The first example is also a case of digital representation, although the second is not: more on this distinction is given in the next section.

17.   9.我使用限定词“在某种意义上”，因为尽管实数是连续的，但单个实数却不是连续的，而且实数的所有数字展开式的集合本身也不是连续的。

18.   9.  I use the qualifier “in a sense” because even though the real numbers are continuous, a single real number is not, and it is not the case that the set of all digital expansions of real numbers is itself continuous.

19. 10。对于高于 10 的基数，将使用新符号作为数字。十六进制（基数为 16）系统使用 A 表示 10，B 表示 11，依此类推，因此“A2F” = ( A ×16 ² ) + (2 ×16 ¹ ) + ( F ×16 ⁰ ) = (10 ×256) + (2 ×16) + (15 ×1) = 2607。

20. 10.  In the case of bases higher than 10, new symbols are used as numerals. The hexadecimal (base-16) system uses A for 10, B for 11, and so on, so that “A2F” = (A×16²) + (2 ×16¹) + (F ×16⁰) = (10 ×256) + (2 ×16) + (15 ×1) = 2607.

21. 11.例如，48 π被解释为一个两位数（四十八）与一个符号表示的一（π）的乘积，但不能解释为一个三位数。

22. 11.  For example, 48π is interpreted as the multiplication of a two-digit number (forty-eight) with a symbolically represented one (π), but not as a three-digit number.

23. 12.通常，“1”用5伏表示，“0”用0伏表示，但也有其他惯例。

24. 12.  Usually, “1” is represented by five volts, and “0” is represented by zero volts, but there are other conventions.

第二部分：什么是智力？

Part II: What Is Intelligence?

人工智能 (AI) 的核心目标是制造智能机器。但​​机器“智能”是什么意思？我们如何判断机器是否真正智能？本部分的文章对这些基本问题持不同立场。

The central aim of artificial intelligence (AI) is to build intelligent machines. But what does it mean for a machine to be”intelligent”? And how can we tell when a machine is, in fact, intelligent? The articles in this part take different stances on these fundamental questions.

根据一种主流传统，追溯到图灵的《计算机器与智能》（第 6 章），一个答案足以回答这两个问题。在那篇文章中，图灵介绍了电影中用于智能的“图灵测试”，该测试使用人类对话范式作为评估机器的试金石。该测试将“智能机器”操作化为一种可以让观察者相信它（很可能）是一个（典型）人类的机器。正如法官波特·斯图尔特对色情的定义一样，我们看到智能时就知道它是什么。此外，我们通过将其与典型进行比较来了解它：在人工智能的情况下，与人类智能行为进行比较。

According to one dominant tradition, tracing to Turing’s “Computing Machinery and Intelligence” (chapter 6), one answer suffices for both questions. In that article, Turing introduces the movie-worthy “Turing Test” for intelligence, which uses the paradigm of human conversation as the touchstone for assessing machines. That test operationalizes “intelligent machine” as a machine that can convince observers more often than not that it is (likely) a (typical) human. As with Justice Potter Stewart’s definition of pornography, we know intelligence when we see it. Furthermore, we come to know it by comparing it to exemplars: in the case of AI, to human intelligent behavior.

Leveseque 的论点（第 7 章）更为狭隘和具有约束力：即使承认符号操作足以理解，图灵测试也太容易被操纵了。第一批聊天机器人之一 Eliza 扮演了罗杰斯心理治疗师的角色，只需将问题转向对话者：“你确定吗？”“那你感觉如何？”“你感觉那样多久了？”“我不确定我是否完全理解你。”尽管 Eliza 很难说服坚定的对话者，但它展示了一种未来聊天机器人更有效地模仿的策略：使用幽默和其他转移话题的策略来创造一个迷人的谈话者，善于分散对话者的注意力，使他们不去评估对话者是否真的知道发生了什么。Levesque 声称，所需要的是对计算机是否理解以及理解什么进行更直接的测试。为此，他认为，“Winograd 图式”可用于测试对话者在常见背景知识能够清楚、果断地解决难题的情况下，运用背景知识消除重复代词指称歧义的能力。

Leveseque’s argument (in chapter 7) is narrower and more proscriptive: even granting that symbol manipulation might suffice for understanding, the Turing test is too easily gamed. One of the first chatbots, Eliza, plays the role of a Rogerian psychotherapist by simply turning questions around on the interlocutor: “Are you sure?” “And how does [that] make you feel?” “How long have you been [feeling like that]?” “I’m not sure I understand you fully.” Although Eliza is hardly convincing to a determined interlocutor, it shows a kind of strategy that future chatbots emulate more effectively: using humor and other deflection strategies to create a charming conversationalist, good at distracting the interlocutor from assessing whether its conversation partner really knows what’s going on. What is required, Levesque claims, is a more direct test of whether and what the computer understands. For this, he argues that “Winograd schemas” can be used to test an interlocutor’s ability to use background knowledge to disambiguate anaphoric pronoun reference in cases in which common background knowledge clearly and decisively solves the puzzle.

在她那篇内容广泛、发人深省的文章（第 10 章）中，梅兰妮·米切尔 (Melanie Mitchell) 认为，我们还没有对智能有一个清晰的概念。她声称，这种缺乏指导思想至少是造成人工智能历史上出现热情（人工智能之春）和失望（人工智能之冬）周期的部分原因。她指出了四个基本的“谬误”，她认为这些谬误让人工智能研究人员产生了一种人为的感觉，认为他们正在朝着通用人工智能的目标前进——也就是说，朝着或多或少能捕捉到典型的有能力的成年人如何运作的人工智能前进。所有这些谬误中的一个统一主题是，狭隘的、特定领域的解决问题能力（例如，人脸识别）与广泛的、领域通用的知识（似乎毫不费力地进入人类大脑）之间存在差异。

In her broad and provocative essay (chapter 10), Melanie Mitchell argues that we do not yet have a clear conception of what intelligence is. She claims that that this lack of a guiding idea is at least partially to blame for the cycles of enthusiasm (AI springs) and disappointment (AI winters) that characterize the history of AI. She identifies four fundamental “fallacies” that, she argues, give AI researchers an artificial sense of progress toward the goal of generalized AI—that is, toward AI that more or less captures how typical competent adults function. One unifying theme in all these fallacies is a difference between the kind of narrow, domain-specific problem-solving capacity that produces so much excitement (e.g., face recognition) and the kind of broad, domain-general knowledge that seems to come so effortlessly to the human mind.

领域特定能力和领域通用能力之间的这种关系正是福多尔经典的垂直（狭义、领域特定、信息封装的处理器）和水平（广义、综合、信息多孔）系统之间的区别。福多尔的核心论点（第 9 章）是，领域特定处理系统只能让我们在一定程度上理解思维。我们认可信念的能力（如科学信念的发展）、我们使用语言描述领域特定感官输入的能力，以及以有目的的、目标驱动的方式整合感知和行动的能力，似乎都需要获得无法事先指定的广泛、通用的信息。福多尔在此基础上认为，一定存在非模块化系统（即信息封装不狭义的系统），它们可以整合来自不同模块化系统的输入，以实现人类典型的信念形成和代理能力。

This relationship between domain-specific and domain-general capacities is precisely what is at stake in Fodor’s classic distinction between vertical (narrow, domain-specific, informationally encapsulated processers) and horizontal (broad, integrative, informationally porous) systems. Fodor’s central thesis (chapter 9) is that domain-specific processing systems can take us only so far in our understanding of the mind. Our ability to endorse beliefs (as in the development of scientific beliefs), our ability to use language to describe domain-specific sensory inputs, and the integration of perception and action in purposive, goal-driven ways all seem to require access to broad, general information that one cannot specify in advance. Fodor argues, on this basis, that there must be non-modular systems (i.e., systems that are not narrowly informationally encapsulated) that can integrate the input from different modular systems to achieve the kind of belief-formation and agential capacities typical of humans.

描述福多尔所追求的那种中枢处理的一种常见方式是使用“理性”一词：人们可能会认为，信念固着和有目的的行为最终可以用一种理性反应来描述，因此，是对正确或不正确推理规则的响应。罗素的文章以人工智能的核心目标是制造理性的机器作为其基本假设。但他问道，理性是什么？它是福多尔所想的那种理性吗？此外，我们希望我们的人工智能机器有多理性？它们的理性必须是完美的（和斯波克一样），还是仅仅达到人类（即明显次优的）理性水平就足够了？此外，理性必须仅体现在行为上，还是内部推理过程也必须模仿人类的推理思维？通过探索这些问题的轮廓，罗素指出了计算系统以及思维的一些明显局限性。例如，时间限制很重要。即使是典型智能的人类也无法考虑所有可能的信息。但这些机器必须工作多快？它们必须有多理性？在努力打造理性的机器时，什么样的理性才构成了适当的目标？在此过程中，罗素探索了人工智能问题的“难度”程度，并提出了有限理性理论，他认为，这足以作为该领域的合理目标。

One common way to characterize the kind of central processing that Fodor is after is in terms of “rationality”: belief fixation and purposive action are, one might think, definitively characterized in terms of a kind of reason-responsiveness and, as such, a responsiveness to the rules of correct or incorrect inference. Russell’s article takes as its fundamental assumption that a central goal of AI is to build machines that are rational. But, he asks, what is rationality? Is it the kind of rationality Fodor has in mind? Furthermore, how rational do we expect our AI machines to be? Must their rationality be perfect (and Spock-like), or is it sufficient merely to achieve human (i.e., demonstrably sub-optimal) levels of rationality? Additionally, must the rationality be evident in behavior alone, or must the internal reasoning process also ape human ratiocinative thought? By exploring the contours of these questions, Russell notes some of the obvious limitations on computational systems, as well as on minds. Time constraints matter, for example. Even paradigmatically intelligent humans can’t consider all possible information. But how fast must these machines work? And how rational must they be? And what is the sense of rationality that constitutes the appropriate target in the effort to build suitably rational machines? In the process, Russell explores degrees of “hardness” in AI problems and offers a theory of bounded rationality that, he argues, suffices as a reasonable target for the field.

6.1 模仿游戏

6.1  The imitation game

我打算考虑“机器能思考吗？”这个问题。首先应该定义“机器”和“思考”的含义。这些定义可以尽可能反映这些词的正常用法，但这种态度是危险的。如果要通过研究“机器”和“思考”这两个词的常用用法来发现它们的含义，就很难不得出这样的结论：“机器能思考吗？”这个问题的含义和答案应该在盖洛普民意调查等统计调查中寻找。但这是荒谬的。我不会尝试这样的定义，而是用另一个与之密切相关、用相对明确的词语表达的问题来代替这个问题。

I propose to consider the question “Can machines think?” This should begin with definitions of the meaning of the terms ‘machine’ and ‘think’. The definitions might be framed so as to reflect so far as possible the normal use of the words, but this attitude is dangerous. If the meaning of the words ‘machine’ and ‘think’ are to be found by examining how they are commonly used it is difficult to escape the conclusion that the meaning and the answer to the question, “Can machines think?” is to be sought in a statistical survey such as a Gallup poll. But this is absurd. Instead of attempting such a definition I shall replace the question by another, which is closely related to it and is expressed in relatively unambiguous words.

问题的新形式可以用我们称之为“模仿游戏”的游戏来描述。游戏由三个人玩，一个男人（A）、一个女人（B）和一个询问者（C），询问者可以是任何性别。询问者待在一个与其他两个人分开的房间里。询问者的游戏目标是确定其他两个人中哪个是男人，哪个是女人。他通过标签 X 和 Y 知道他们，并在游戏结束时说“X 是 A 而 Y 是 B”或“X 是 B 而 Y 是 A”。询问者可以这样向 A 和 B 提问：

The new form of the problem can be described in terms of a game which we call the “imitation game”. It is played with three people, a man (A), a woman (B), and an interrogator (C) who may be of either sex. The interrogator stays in a room apart from the other two. The object of the game for the interrogator is to determine which of the other two is the man and which is the woman. He knows them by labels X and Y, and at the end of the game he says either “X is A and Y is B” or “X is B and Y is A”. The interrogator is allowed to put questions to A and B thus:

C：X能告诉我他或她的头发长度吗？

C: Will X please tell me the length of his or her hair?

现在假设 X 实际上是 A，那么 A 必须回答。游戏中 A 的目标是试图让 C 做出错误的判断。因此他的答案可能是：

Now suppose X is actually A, then A must answer. It is A’s object in the game to try to cause C to make the wrong identification. His answer might therefore be:

答：我的头发是瓦状的，最长的头发大约有九英寸长。

A: My hair is shingled, and the longest strands are about nine inches long.

为了避免语调对审讯者造成影响，答案应该用书面形式，或者最好用打字机打字。理想的安排是使用电传打字机在两个房间之间进行通信。或者，也可以由中间人重复问题和答案。第三个玩家（B）的游戏目标是帮助审讯者。对她来说，最好的策略可能是给出真实的答案。她可以在答案中添加诸如“我是女人，别听他的！”之类的话，但这毫无用处，因为男人也可以做出类似的评论。

In order that tones of voice may not help the interrogator the answers should be written, or better still, typewritten. The ideal arrangement is to have a teleprinter communicating between the two rooms. Alternatively the question and answers can be repeated by an intermediary. The object of the game for the third player (B) is to help the interrogator. The best strategy for her is probably to give truthful answers. She can add such things as “I am the woman, don’t listen to him!” to her answers, but it will avail nothing as the man can make similar remarks.

我们现在要问的是：“如果一台机器在这场游戏中扮演 A 的角色，会发生什么？”当游戏以这种方式进行时，询问者做出错误决定的几率会不会像在男人和女人之间进行游戏时一样高？这些问题取代了我们最初的问题：“机器能思考吗？”

We now ask the question, “What will happen when a machine takes the part of A in this game?” Will the interrogator decide wrongly as often when the game is played like this as he does when the game is played between a man and a woman? These questions replace our original, “Can machines think?”

6.2 对新问题的批判

6.2  Critique of the new problem

除了问“这个新形式的问题的答案是什么？”，我们还可以问“这个新问题值得研究吗？”我们不再多说，直接研究后一个问题，从而缩短了无限倒退的时间。

As well as asking, “What is the answer to this new form of the question?” one may ask, “Is this new question a worthy one to investigate?” This latter question we investigate without further ado, thereby cutting short an infinite regress.

新问题的优点在于，它相当清晰地区分了人类的体能和智力。没有工程师或化学家声称能够生产出与人类皮肤无法区分的材料。也许有一天可以做到这一点，但即使假设这项发明可用，我们也应该觉得，试图通过给“思考机器”穿上这种人造皮肤来使其更像人类是没有意义的。我们设置问题的形式反映了这一事实，即询问者无法看到或触摸其他竞争对手，也无法听到他们的声音。建议标准的其他一些优点可以通过样本问题和答案来显示出来。因此：

The new problem has the advantage of drawing a fairly sharp line between the physical and the intellectual capacities of a man. No engineer or chemist claims to be able to produce a material which is indistinguishable from the human skin. It is possible that at some time this might be done, but even supposing this invention available we should feel there was little point in trying to make a “thinking machine” more human by dressing it up in such artificial flesh. The form in which we have set the problem reflects this fact in the condition which prevents the interrogator from seeing or touching the other competitors, or hearing their voices. Some other advantages of the proposed criterion may be shown up by specimen questions and answers. Thus:

问答法似乎适用于介绍我们希望涵盖的几乎任何一个人类活动领域。我们不想因为机器在选美比赛中表现不佳而惩罚它，也不想因为人类在与飞机的比赛中失败而惩罚它。我们的游戏条件使这些缺陷变得无关紧要。如果“目击者”认为合适，他们可以随心所欲地吹嘘自己的魅力、力量或英雄气概，但询问者不能要求他们进行实际演示。

The question and answer method seems to be suitable for introducing almost any one of the fields of human endeavor that we wish to include. We do not wish to penalize the machine for its inability to shine in beauty competitions, nor to penalize a man for losing in a race against an airplane. The conditions of our game make these disabilities irrelevant. The “witnesses” can brag, if they consider it advisable, as much as they please about their charms, strength or heroism, but the interrogator cannot demand practical demonstrations.

也许有人会批评这个游戏，认为机器获胜的几率太低了。如果人试图假装自己是机器，他的表现显然会很糟糕。他会因为算术缓慢和不准确而立刻暴露自己。机器难道不能做一些应该被描述为思考的事情，但这与人做的事情截然不同吗？这个反对意见非常强烈，但至少我们可以说，如果一台机器能够令人满意地玩模仿游戏，我们就不必为这个反对意见而烦恼。

The game may perhaps be criticized on the ground that the odds are weighted too heavily against the machine. If the man were to try and pretend to be the machine he would clearly make a very poor showing. He would be given away at once by slowness and inaccuracy in arithmetic. May not machines carry out something which ought to be described as thinking but which is very different from what a man does? This objection is a very strong one, but at least we can say that if, nevertheless, a machine can be constructed to play the imitation game satisfactorily, we need not be troubled by this objection.

有人可能会说，在玩“模仿游戏”时，机器的最佳策略可能不是模仿人类的行为。这可能是真的，但我认为这种策略不太可能产生任何重大影响。无论如何，这里无意研究游戏理论，我们将假设最佳策略是尝试提供人类自然会给出的答案。

It might be urged that when playing the “imitation game” the best strategy for the machine may possibly be something other than imitation of the behavior of a man. This may be, but I think it is unlikely that there is any great effect of this kind. In any case there is no intention to investigate here the theory of the game, and it will be assumed that the best strategy is to try to provide answers that would naturally be given by a man.

6.3 参与游戏的机器

6.3  The machines concerned in the game

在我们明确“机器”一词的含义之前，我们在第 6.1 节中提出的问题将不会十分明确。我们自然希望允许在我们的机器中使用每一种工程技术。我们还希望允许这样一种可能性，即工程师或工程师团队可以制造一台可以工作的机器，但其操作方式无法由其构造者令人满意地描述，因为他们采用的方法主要是实验性的。最后，我们希望将以通常方式出生的男性排除在机器之外。很难制定满足这三个条件的定义。例如，人们可能会坚持认为工程师团队应该全部是同一种性别，但这实际上并不令人满意，因为可能有可能从一个人皮肤的一个细胞（比如说）中培养出一个完整的个体。这样做将是一项值得最高赞扬的生物技术壮举，但我们不会倾向于将其视为“建造一台思考机器”的案例。这促使我们放弃允许使用每种技术的要求。鉴于目前人们对“思维机器”的兴趣是由一种特殊的机器引起的，这种机器通常被称为“电子计算机”或“数字计算机”，因此我们更愿意这样做。根据这一建议，我们只允许数字计算机参与我们的游戏。

The question which we put in Section 6.1 will not be quite definite until we have specified what we mean by the word ‘machine’. It is natural that we should wish to permit every kind of engineering technique to be used in our machines. We also wish to allow the possibility that an engineer or team of engineers may construct a machine which works, but whose manner of operation cannot be satisfactorily described by its constructors because they have applied a method which is largely experimental. Finally, we wish to exclude from the machines men born in the usual manner. It is difficult to frame the definitions so as to satisfy these three conditions. One might for instance insist that the team of engineers should be all of one sex, but this would not really be satisfactory, for it is probably possible to rear a complete individual from a single cell of the skin (say) of a man. To do so would be a feat of biological technique deserving of the very highest praise, but we would not be inclined to regard it as a case of “constructing a thinking machine”. This prompts us to abandon the requirement that every kind of technique should be permitted. We are the more ready to do so in view of the fact that the present interest in “thinking machines” has been aroused by a particular kind of machine, usually called an “electronic computer” or “digital computer”. Following this suggestion we only permit digital computers to take part in our game.

乍一看，这个限制似乎非常严格。我将尝试证明事实并非如此。要做到这一点，需要对这些计算机的性质和属性进行简要说明。

This restriction appears at first sight to be a very drastic one. I shall attempt to show that it is not so in reality. To do this necessitates a short account of the nature and properties of these computers.

也可以说，将机器等同于数字计算机的做法，就像我们的“思考”标准一样，只有在数字计算机无法在游戏中表现良好时（与我的看法相反）才会令人不满意。

It may also be said that this identification of machines with digital computers, like our criterion for “thinking”, will only be unsatisfactory if (contrary to my belief), it turns out that digital computers are unable to give a good showing in the game.

目前已经有许多数字计算机投入使用，人们可能会问：“为什么不立即进行实验呢？满足游戏条件很容易。可以使用许多询问器，并编制统计数据以显示正确识别的频率。”简而言之，我们不是在问所有数字计算机是否都能在游戏中表现出色，也不是问目前可用的计算机是否表现出色，而是问是否有可以想象的计算机会表现出色。但这只是简短的回答。我们稍后会从不同的角度看待这个问题。

There are already a number of digital computers in working order, and it may be asked, “Why not try the experiment straight away? It would be easy to satisfy the conditions of the game. A number of interrogators could be used, and statistics compiled to show how often the right identification was given.” The short answer is that we are not asking whether all digital computers would do well in the game nor whether the computers at present available would do well, but whether there are imaginable computers which would do well. But this is only the short answer. We shall see this question in a different light later.

6.4 数字计算机

6.4  Digital computers

数字计算机背后的理念可以这样解释：这些机器旨在执行人类计算机可以完成的任何操作。人类计算机应该遵循固定的规则；他无权在任何细节上偏离这些规则。我们可以假设这些规则写在一本书中，每当他被安排到新的工作时，这本书就会被修改。他还有无限量的纸张来做他的计算。他也可以在“台式机”上做乘法和加法，但这并不重要。

The idea behind digital computers may be explained by saying that these machines are intended to carry out any operations which could be done by a human computer. The human computer is supposed to be following fixed rules; he has no authority to deviate from them in any detail. We may suppose that these rules are supplied in a book, which is altered whenever he is put on to a new job. He has also an unlimited supply of paper on which he does his calculations. He may also do his multiplications and additions on a “desk machine”, but this is not important.

如果我们用上述解释作为定义，就会陷入循环论证的危险。我们可以通过概述实现预期效果的方法避免这种情况。数字计算机通常可视为由三部分组成：

If we use the above explanation as a definition, we shall be in danger of circularity of argument. We avoid this by giving an outline of the means by which the desired effect is achieved. A digital computer can usually be regarded as consisting of three parts:

存储器是信息的存储器，对应于人类计算机的纸张，无论这是他进行计算的纸张还是打印规则书的纸张。只要人类计算机在头脑中进行计算，存储器的一部分将对应于他的记忆。

The store is a store of information, and corresponds to the human computer’s paper, whether this is the paper on which he does his calculations or that on which his book of rules is printed. Insofar as the human computer does calculations in his head, a part of the store will correspond to his memory.

执行单元是执行计算中涉及的各种单个操作的部分。这些单个操作因机器而异。通常可以执行相当长的运算，例如“将 3540675445 乘以 7076345687”，但在某些机器中只能执行非常简单的运算，例如“写下 0”。

The executive unit is the part which carries out the various individual operations involved in a calculation. What these individual operations are will vary from machine to machine. Usually fairly lengthy operations, such as “Multiply 3540675445 by 7076345687”, can be done, but in some machines only very simple ones, such as “Write down 0”, are possible.

我们提到过，计算机中提供的“规则手册”在机器中被存储器的一部分所取代。它被称为“指令表”。控制的职责是确保这些指令得到正确且按正确的顺序执行。控制的构造必然会发生这种情况。

We have mentioned that the “book of rules” supplied to the computer is replaced in the machine by a part of the store. It is then called the “table of instructions”. It is the duty of the control to see that these instructions are obeyed correctly and in the right order. The control is so constructed that this necessarily happens.

存储器中的信息通常被分成相当小的数据包。例如，在一台机器中，一个数据包可能由十个十进制数字组成。以某种系统方式将数字分配给存储器中存储各种信息包的部分。典型的指令可能是：

The information in the store is usually broken up into packets of moderately small size. In one machine, for instance, a packet might consist of ten decimal digits. Numbers are assigned to the parts of the store in which the various packets of information are stored, in some systematic manner. A typical instruction might say:

不用说，它不会出现在用英语表达的机器中。它更可能被编码成诸如 6809430217 的形式。这里 17 表示要对两个数字执行各种可能的操作中的哪一种——在这种情况下是上面描述的操作，即“添加数字…… ”。会注意到该指令占用 10 位数字，因此非常方便地形成了一个信息包。控制器通常会按照指令存储位置的顺序执行要执行的指令，但偶尔会出现这样的指令

Needless to say it would not occur in the machine expressed in English. It would more likely be coded in a form such as 6809430217. Here 17 says which of various possible operations is to be performed on the two numbers—in this case the operation that is described above, namely, “Add the number…”. It will be noticed that the instruction takes up 10 digits and so forms one packet of information, very conveniently. The control will normally take the instructions to be obeyed in the order of the positions in which they are stored, but occasionally an instruction such as

可能会遇到，或者再次

may be encountered, or again

后一种类型的指令非常重要，因为它们使得一系列操作可以一遍又一遍地重复，直到满足某个条件为止，但这样做时，要遵守的不是每次重复的新指令，而是一遍又一遍地重复相同的指令。举一个家庭的类比，假设母亲希望汤米每天早上在上学的路上去鞋匠那里看看她的鞋子是否修好了。她可以每天早上重新问他。或者，她可以一劳永逸地在走廊里贴一张通知，汤米上学时会看到这张通知，上面告诉他去取鞋子，如果他带着鞋子回来，还要把这张通知毁掉。

Instructions of these latter types are very important because they make it possible for a sequence of operations to be repeated over and over again until some condition is fulfilled, but in doing so to obey, not fresh instructions on each repetition, but the same ones over and over again. To take a domestic analogy, suppose Mother wants Tommy to call at the cobbler’s every morning on his way to school to see if her shoes are done. She can ask him afresh every morning. Alternatively she can stick up a notice once and for all in the hall which he will see when he leaves for school and which tells him to call for the shoes, and also to destroy the notice when he comes back if he has the shoes with him.

读者必须接受这样一个事实：根据我们所描述的原理，数字计算机是可以构建的，而且也确实已经构建好了，而且它们实际上可以非常接近地模仿人类计算机的行为。

The reader must accept it as a fact that digital computers can be constructed, and indeed have been constructed, according to the principles we have described, and that they can in fact mimic the actions of a human computer very closely.

我们描述的人类计算机所使用的规则书当然是一种方便的虚构。实际的人类计算机确实记得它们必须做什么。如果想让机器模仿人类计算机在某些复杂操作中的行为，就必须问它是如何完成的，然后将答案转换成指令表的形式。构建指令表通常被描述为“编程”。对机器进行编程以执行操作 A 意味着将适当的指令表放入机器中，以便它执行 A。

The book of rules which we have described our human computer as using is of course a convenient fiction. Actual human computers really remember what they have got to do. If one wants to make a machine mimic the behavior of the human computer in some complex operation one has to ask him how it is done, and then translate the answer into the form of an instruction table. Constructing instruction tables is usually described as “programming”. To “program a machine to carry out the operation A” means to put the appropriate instruction table into the machine so that it will do A.

数字计算机概念的一个有趣变体是带有随机元素的数字计算机。这些计算机具有涉及掷骰子或某种等效电子过程的指令；例如，其中一条指令可能是

An interesting variant on the idea of a digital computer is a digital computer with a random element. These have instructions involving the throwing of a die or some equivalent electronic process; one such instruction might for instance be

有时，这样的机器被描述为具有自由意志（尽管我自己不会使用这个短语）。通常不可能通过观察机器来确定它是否具有随机元素，因为这种设备可以产生类似的效果，例如根据π的小数位数做出选择。

Sometimes such a machine is described as having free will (though I would not use this phrase myself). It is not normally possible to determine from observing a machine whether it has a random element, for a similar effect can be produced by such devices as making the choices depend on the digits of the decimal for π.

大多数实际数字计算机只有有限的存储空间。拥有无限存储空间的计算机的概念在理论上没有任何困难。当然，一次只能使用其中的有限部分。同样，只能构建有限数量的存储空间，但我们可以想象根据需要添加越来越多的存储空间。这样的计算机具有特殊的理论意义，被称为无限容量计算机。

Most actual digital computers have only a finite store. There is no theoretical difficulty in the idea of a computer with an unlimited store. Of course only a finite part of it can have been used at any one time. Likewise only a finite amount can have been constructed, but we can imagine more and more being added as required. Such computers have special theoretical interest and will be called infinite capacity computers.

数字计算机的概念由来已久。查尔斯·巴贝奇（Charles Babbage）曾于 1828 年至 1839 年担任剑桥大学卢卡斯数学教授，他曾设计过这样一种机器，称为“分析机”，但从未完成。尽管巴贝奇掌握了所有基本概念，但他的机器当时并不那么有吸引力。当时的速度肯定比人类计算机快，但比曼彻斯特机慢 100 倍左右，而曼彻斯特机本身是现代机器中速度较慢的机器之一。存储是纯机械的，使用轮子和卡片。

The idea of a digital computer is an old one. Charles Babbage, Lucasian Professor of Mathematics at Cambridge from 1828 to 1839, planned such a machine, called the “Analytical Engine”, but it was never completed. Although Babbage had all the essential ideas, his machine was not at that time such a very attractive prospect. The speed which would have been available would be definitely faster than a human computer but something like 100 times slower than the Manchester machine, itself one of the slower of the modern machines. The storage was to be purely mechanical, using wheels and cards.

巴贝奇的分析机完全是机械的，这一事实将帮助我们摆脱迷信。人们常常重视现代数字计算机是电子的，而神经系统也是电子的。由于巴贝奇的机器不是电子的，而且所有数字计算机在某种意义上都是等效的，因此，我们看到这种电的使用在理论上并不重要。当然，电通常用于快速信号传输，因此我们在这两种情况下都发现电也就不足为奇了。在神经系统中，化学现象至少与电一样重要。在某些计算机中，存储系统主要是声学的。因此，使用电的特性被视为只是一种非常表面的相似性。如果我们想找到这种相似性，我们应该寻找函数的数学类比。

The fact that Babbage’s Analytical Engine was to be entirely mechanical will help us to rid ourselves of a superstition. Importance is often attached to the fact that modern digital computers are electrical, and that the nervous system also is electrical. Since Babbage’s machine was not electrical, and since all digital computers are in a sense equivalent, we see that this use of electricity cannot be of theoretical importance. Of course electricity usually comes in where fast signaling is concerned, so it is not surprising that we find it in both these connections. In the nervous system chemical phenomena are at least as important as electrical. In certain computers the storage system is mainly acoustic. The feature of using electricity is thus seen to be only a very superficial similarity. If we wish to find such similarities we should look rather for mathematical analogies of function.

6.5 数字计算机的通用性

6.5  Universality of digital computers

上一节中讨论的数字计算机可以归类为“离散状态机”。这些机器通过突然跳跃或点击从一个非常明确的状态移动到另一个状态。这些状态足够不同，以至于可以忽略它们之间混淆的可能性。严格地说，没有这样的机器。一切都在连续移动。但是有许多类型的机器可以有利地被认为是离散状态机。例如，在考虑照明系统的开关时，每个开关必须明确打开或明确关闭是一种方便的虚构。必须有中间位置，但对于大多数目的，我们可以忘记它们。作为离散状态机的一个例子，我们可以考虑一个每秒点击 120 ^°的轮子，但可以通过可以从外部操作的杠杆将其停止；此外，灯将在轮子的某个位置点亮。该机器可以抽象地描述如下：机器的内部状态（由轮子的位置描述）可能是q ₁、q ₂或q ₃。有一个输入信号i ₀或i ₁（杠杆位置）。任何时刻的内部状态由最后一个状态和输入信号根据表格确定

The digital computers considered in the last section may be classified among the “discrete state machines”. These are the machines which move by sudden jumps or clicks from one quite definite state to another. These states are sufficiently different for the possibility of confusion between them to be ignored. Strictly speaking there are no such machines. Everything really moves continuously. But there are many kinds of machines which can profitably be thought of as being discrete state machines. For instance in considering the switches for a lighting system it is a convenient fiction that each switch must be definitely on or definitely off. There must be intermediate positions, but for most purposes we can forget about them. As an example of a discrete state machine, we might consider a wheel which clicks round through 120^° once a second, but may be stopped by a lever which can be operated from outside; in addition a lamp is to light in one of the positions of the wheel. This machine could be described abstractly as follows: The internal state of the machine (which is described by the position of the wheel) may be q₁, q₂, or q₃. There is an input signal i₀ or i₁ (position of lever). The internal state at any moment is determined by the last state and input signal according to the table

输出信号是内部状态（光）的唯一外部可见指示，如表所示

The output signals, the only externally visible indication of the internal state (the light), are described by the table

这个例子是典型的离散状态机。它们可以用这样的表格来描述，只要它们只有有限个可能的状态。

This example is typical of discrete state machines. They can be described by such tables, provided they have only a finite number of possible states.

似乎只要给定机器的初始状态和输入信号，就总是可以预测所有未来状态。这让人想起拉普拉斯的观点，即从某一时刻宇宙的完整状态（由所有粒子的位置和速度描述）出发，应该可以预测所有未来状态。然而，我们正在考虑的预测比拉普拉斯考虑的预测更接近实用性。“整个宇宙”的系统是这样的，初始条件中非常小的误差可能会在以后产生压倒性的影响。单个电子在某一时刻位移十亿分之一厘米可能会导致一个人一年后被雪崩杀死或逃脱。我们称之为“离散状态机”的机械系统的一个基本特性是这种现象不会发生。即使我们考虑实际的物理机器而不是理想化的机器，对某一时刻状态的合理准确了解也会在以后的任何步骤中产生合理准确的了解。

It will seem that given the initial state of the machine and the input signals it is always possible to predict all future states. This is reminiscent of Laplace’s view that from the complete state of the universe at one moment of time, as described by the positions and velocities of all particles, it should be possible to predict all future states. The prediction which we are considering is, however, rather nearer to practicability than that considered by Laplace. The system of the “universe as a whole” is such that quite small errors in the initial conditions can have an overwhelming effect at a later time. The displacement of a single electron by a billionth of a centimeter at one moment might make the difference between a man being killed by an avalanche a year later, or escaping. It is an essential property of the mechanical systems which we have called “discrete state machines” that this phenomenon does not occur. Even when we consider the actual physical machines instead of the idealized machines, reasonably accurate knowledge of the state at one moment yields reasonably accurate knowledge any number of steps later.

我们已经提到过，数字计算机属于离散状态机。但是这种机器能够达到的状态数通常非常大。例如，现在在曼彻斯特工作的机器的状态数约为 2 ^165,000 —即约 10 ^50,000。将其与上面描述的具有三个状态的打孔轮示例进行比较。不难看出为什么状态数应该如此巨大。计算机包括一个与人类计算机使用的纸张相对应的存储器。必须能够将可能写在纸上的任意一种符号组合写入存储器。为简单起见，假设仅使用从 0 到 9 的数字作为符号。忽略笔迹的变化。假设允许计算机100 张纸，每张纸包含 50 行，每行可容纳 30 位数字。那么状态数就是 10 ^100×50×30 — 也就是 10 ^150,000。这大约是三台曼彻斯特机加在一起的状态数。状态数以 2 为底的对数通常称为机器的“存储容量”。因此，曼彻斯特机的存储容量大约为 165,000，而我们示例中的轮机的存储容量大约为 1.6。如果将两台机器放在一起，必须将它们的容量相加才能得到合成机器的容量。这就导致可能出现这样的陈述：“曼彻斯特机包含 64 个磁道，每个磁道的容量为 2560，8 个电子管，容量为 1280。杂项存储空间大约为 300，总计 174,380。”

As we have mentioned, digital computers fall within the class of discrete state machines. But the number of states of which such a machine is capable is usually enormously large. For instance, the number for the machine now working at Manchester is about 2^{165, 000}—that is, about 10^{50, 000}. Compare this with our example of the dicking wheel described above, which had three states. It is not difficult to see why the number of states should be so immense. The computer includes a store corresponding to the paper used by a human computer. It must be possible to write into the store any one of the combinations of symbols which might have been written on the paper. For simplicity suppose that only digits from 0 to 9 are used as symbols. Variations in handwriting are ignored. Suppose the computer is allowed 100 sheets of paper each containing 50 lines each with room for 30 digits. Then the number of states is 10^100×50×30—that is, 10^{150, 000}. This is about the number of states of three Manchester machines put together. The logarithm to the base two of the number of states is usually called the “storage capacity” of the machine. Thus the Manchester machine has a storage capacity of about 165,000 and the wheel machine of our example about 1.6. If two machines are put together their capacities must be added to obtain the capacity of the resultant machine. This leads to the possibility of statements such as “The Manchester machine contains 64 magnetic tracks each with a capacity of 2560, eight electronic tubes with a capacity of 1280. Miscellaneous storage amounts to about 300 making a total of 174,380.”

给定与离散状态机相对应的表，就可以预测它将做什么。没有理由不通过数字计算机进行这种计算。只要能够足够快地完成，数字计算机就可以模仿任何离散状态机的行为。然后可以用所讨论的机器（作为 B）和模仿的数字计算机（作为 A）玩模仿游戏，询问者将无法区分它们。当然，数字计算机必须具有足够的存储容量，并且工作速度足够快。此外，它必须为想要模仿的每台新机器重新编程。

Given the table corresponding to a discrete state machine, it is possible to predict what it will do. There is no reason why this calculation should not be carried out by means of a digital computer. Provided it could be carried out sufficiently quickly the digital computer could mimic the behavior of any discrete state machine. The imitation game could then be played with the machine in question (as B) and the mimicking digital computer (as A) and the interrogator would be unable to distinguish them. Of course the digital computer must have adequate storage capacity a well as working sufficiently fast. Moreover, it must be programmed afresh for each new machine which it is desired to mimic.

数字计算机的这一特殊属性，即它们可以模拟任何离散状态机，被称为通用机器。具有这种属性的机器的存在具有重要的意义，即，除了速度之外，无需设计各种新机器来执行各种计算过程。它们都可以用一台数字计算机完成，并针对每种情况进行适当的编程。由此可见，所有数字计算机在某种意义上都是等价的。

This special property of digital computers, that they can mimic any discrete state machine, is described by saying that they are universal machines. The existence of machines with this property has the important consequence that, considerations of speed apart, it is unnecessary to design various new machines to do various computing processes. They can all be done with one digital computer, suitably programmed for each case. It will be seen that as a consequence of this all digital computers are in a sense equivalent.

现在我们可以再次考虑第 6.3 节末尾提出的观点。有人试探性地建议将“机器能思考吗？”这个问题替换为“是否存在可以想象的数字计算机能够在模仿游戏中表现良好？”如果我们愿意，我们可以将这个问题从表面上变得更具普遍性，并问：“是否存在表现良好的离散状态机？”但考虑到普遍性，我们发现这两个问题都等同于：“让我们将注意力集中在一台特定的数字计算机 C 上。通过修改这台计算机以使其具有足够的存储空间，适当提高其运行速度，并为其提供适当的程序，C 是否可以令人满意地扮演模仿游戏中 A 的角色，而 B 的角色则由人类扮演？”

We may now consider again the point raised at the end of Section 6.3. It was suggested tentatively that the question, “Can machines think?” should be replaced by “Are there imaginable digital computers which would do well in the imitation game?” If we wish we can make this superficially more general and ask, “Are there discrete state machines which would do well?” But in view of the universality property we see that either of these questions is equivalent to this: “Let us fix our attention on one particular digital computer C. Is it true that by modifying this computer to have an adequate storage, suitably increasing its speed of action, and providing it with an appropriate program, C can be made to play satisfactorily the part of A in the imitation game, the part of B being taken by a man?”

6.6 对主要问题的相反观点

6.6  Contrary views on the main question

现在我们可以认为基础已经清理完毕，我们准备继续讨论我们的问题“机器能思考吗？”以及上一节末尾引用的变体。我们不能完全放弃问题的原始形式，因为对于替换的适当性会有不同意见，我们至少必须听听在这方面该说些什么。

We may now consider the ground to have been cleared and we are ready to proceed to the debate on our question, “Can machines think?” and the variant of it quoted at the end of the last section. We cannot altogether abandon the original form of the problem, for opinions will differ as to the appropriateness of the substitution and we must at least listen to what has to be said in this connection.

如果我先解释一下我自己对这个问题的看法，读者会更容易理解。首先考虑一下这个问题的更准确形式。我相信，大约五十年后，就有可能对存储容量约为 10 9 的计算机进行编程，^使它们能够很好地玩模仿游戏，以至于普通询问者在五分钟的询问后，做出正确识别的几率不会超过 70%。最初的问题“机器能思考吗？”我认为毫无意义，不值得讨论。然而，我相信，到本世纪末，词语的使用和一般受过教育的观点将发生很大变化，人们将能够谈论机器思考，而不会期望遭到反驳。我进一步认为，隐瞒这些信念没有任何用处。普遍的观点认为，科学家们会不可避免地从已证实的事实走向已证实的事实，从不受任何未经证实的猜想的影响，这是完全错误的。只要明确哪些是已证实的事实，哪些是猜想，就不会造成任何伤害。推测非常重要，因为它们可以指明有用的研究方向。

It will simplify matters for the reader if I explain first my own beliefs in the matter. Consider first the more accurate form of the question. I believe that in about fifty years’time it will be possible to program computers, with a storage capacity of about 10⁹, to make them play the imitation game so well that an average interrogator will not have more than 70 per cent chance of making the right identification after five minutes of questioning. The original question, “Can machines think?” I believe to be too meaningless to deserve discussion. Nevertheless I believe that at the end of the century the use of words and general educated opinion will have altered so much that one will be able to speak of machines thinking without expecting to be contradicted. I believe further that no useful purpose is served by concealing these beliefs. The popular view that scientists proceed inexorably from well-established fact to well-established fact, never being influenced by any unproved conjecture, is quite mistaken. Provided it is made clear which are proved facts and which are conjectures, no harm can result. Conjectures are of great importance since they suggest useful lines of research.

我现在开始考虑与我自己的观点相反的观点。

I now proceed to consider opinions opposed to my own.

（1）神学上的反对意见。思考是人类不朽灵魂的功能。上帝赋予每个男人和女人不朽的灵魂，但没有赋予其他动物或机器不朽的灵魂。因此，没有动物或机器能够思考^。1

(1) THE THEOLOGICAL OBJECTION. Thinking is a function of man’s immortal soul. God has given an immortal soul to every man and woman, but not to any other animal or to machines. Hence no animal or machine can think.¹

我无法接受其中的任何部分，但将尝试用神学术语来回答。如果将动物归类为人类，我会发现这个论点更有说服力，因为在我看来，典型的有生命和无生命之间的差异比人类和其他动物之间的差异更大。如果我们考虑一下其他宗教团体成员对正统观点的看法，就会更清楚地看出这种观点的任意性。基督徒如何看待穆斯林认为女性没有灵魂的观点？但让我们把这一点放在一边，回到主要论点上。在我看来，上面引用的论点意味着对全能者的全能的严重限制。人们承认，有些事情他不能做，比如把一等于二，但我们难道不应该相信，如果他认为合适，他有自由赋予大象灵魂吗？我们可以预料，他只会在与突变相结合的情况下行使这种权力，这种突变为大象提供了适当改进的大脑，以满足这个灵魂的需要。对于机器的情况，可以提出完全类似的论点。这看起来可能有所不同，因为它更难“接受”。但这实际上只意味着我们认为他不太可能认为这种情况适合赋予灵魂。本文的其余部分将讨论所讨论的情况。在尝试建造这样的机器时，我们不应该不敬地篡夺他创造灵魂的权力，就像我们在生育孩子方面一样：相反，无论哪种情况，我们都是他意志的工具，为他创造的灵魂提供住所。

I am unable to accept any part of this, but will attempt to reply in theological terms. I should find the argument more convincing if animals were classed with men, for there is a greater difference, to my mind, between the typical animate and the inanimate than there is between man and the other animals. The arbitrary character of the orthodox view becomes clearer if we consider how it might appear to a member of some other religious community. How do Christians regard the Moslem view that women have no souls? But let us leave this point aside and return to the main argument. It appears to me that the argument quoted above implies a serious restriction of the omnipotence of the Almighty. It is admitted that there are certain things that He cannot do such as making one equal to two, but should we not believe that He has freedom to confer a soul on an elephant if He sees fit? We might expect that He would only exercise this power in conjunction with a mutation which provided the elephant with an appropriately improved brain to minister to the needs of this soul. An argument of exactly similar form may be made for the case of machines. It may seem different because it is more difficult to “swallow”. But this really only means that we think it would be less likely that He would consider the circumstances suitable for conferring a soul. The circumstances in question are discussed in the rest of this paper. In attempting to construct such machines we should not be irreverently usurping His power of creating souls, any more than we are in the procreation of children: rather we are, in either case, instruments of His will providing mansions for the souls that He creates.

然而，这只是猜测。我对神学论证并不十分感兴趣，无论它们被用来支持什么。过去，这种论证常常被认为是不令人满意的。在伽利略时代，有人认为“太阳静止不动……整整一天不急于落下”（约书亚记 x. 13）和“他奠定了地球的基础，使它不会在任何时候移动”（诗篇 cv. 5）这些文字足以驳斥哥白尼理论。以我们目前的知识，这样的论点似乎是徒劳的。当没有这些知识时，它给人的印象就大不相同了。

However, this is mere speculation. I am not very impressed with theological arguments, whatever they may be used to support. Such arguments have often been found unsatisfactory in the past. In the time of Galileo it was argued that the texts, “And the sun stood still…and hasted not to go down about a whole day” (Joshua x. 13) and “He laid the foundations of the earth, that it should not move at any time” (Psalm cv. 5) were an adequate refutation of the Copernican theory. With our present knowledge, such an argument appears futile. When that knowledge was not available, it made a quite different impression.

（2）“把头埋在沙子里”的反对意见。“机器思考的后果太可怕了。我们希望并相信它们不会这样做。”

(2) THE “HEADS IN THE SAND” OBJECTION. “The consequences of machines thinking would be too dreadful. Let us hope and believe that they cannot do so.”

这种论点很少像上述形式那样公开表达。但它影响了我们大多数思考它的人。我们喜欢相信人类在某种程度上比其他生物更优越。如果能证明人类必然优越，那就最好了，因为这样他就不会失去他的统治地位。神学论点的流行显然与这种感觉有关。这种感觉在知识分子中可能相当强烈，因为他们比其他人更看重思考的能力，更倾向于将他们对人类优越性的信念建立在这种能力的基础上。

This argument is seldom expressed quite so openly as in the form above. But it affects most of us who think about it at all. We like to believe that Man is in some subtle way superior to the rest of creation. It is best if he can be shown to be necessarily superior, for then there is no danger of him losing his commanding position. The popularity of the theological argument is clearly connected with this feeling. It is likely to be quite strong in intellectual people, since they value the power of thinking more highly than others, and are more inclined to base their belief in the superiority of Man on this power.

我认为这个论点并不足以驳斥。安慰可能更合适；也许应该在灵魂的轮回中寻求安慰。

I do not think that this argument is sufficiently substantial to require refutation. Consolation would be more appropriate; perhaps this should be sought in the transmigration of souls.

(3)数学反对意见。有许多数学逻辑结果可用于表明离散状态机的能力存在限制。这些结果中最著名的是哥德尔定理 (1931)，它表明在任何足够强大的逻辑系统中，都可以制定既不能在系统内证明也不能反驳的陈述，除非系统本身可能不一致。丘奇 (1936)、克莱尼 (1936)、罗瑟 (1936) 和图灵 (1937) 也提出了其他一些在某些方面类似的结果。后一个结果是最方便考虑的，因为它直接涉及机器，而其他结果只能用于相对间接的论证；例如，如果要使用哥德尔定理，我们还需要一些用机器描述逻辑系统的方法，以及用逻辑系统描述机器的方法。所讨论的结果指的是一种本质上是具有无限容量的数字计算机的机器。它指出，有些事情这种机器无法做到。如果像在模仿游戏中一样，它被设计成回答问题，那么对于某些问题，它要么给出错误答案，要么根本无法给出答案，无论回答的时间有多长。当然，可能有很多这样的问题，一台机器无法回答的问题，另一台机器可能会令人满意地回答。当然，我们目前假设这些问题是那种可以用“是”或“否”回答的问题，而不是诸如“你觉得毕加索怎么样？”这样的问题。我们知道机器必定会失败的问题是这种类型，“考虑如下指定的机器……这台机器是否会对任何问题回答‘是’？”这些点将被替换为标准形式的某些机器的描述，该描述可能类似于第 6.5 节中使用的内容。当所描述的机器与正在被询问的机器具有某种相对简单的关系时，可以证明答案是错误的或无法得到的。这是数学结果；有人认为，它证明了机器的缺陷，而人类智力却无法做到这一点。

(3) THE MATHEMATICAL OBJECTION. There are a number of results of mathematical logic which can be used to show that there are limitations to the powers of discrete state machines. The best known of these results is known as Gödel’s theorem (1931), and shows that in any sufficiently powerful logical system statements can be formulated which can neither be proved nor disproved within the system, unless possibly the system itself is inconsistent. There are other, in some respects similar, results due to Church (1936), Kleene (1936), Rosser (1936), and Turing (1937). The latter result is the most convenient to consider, since it refers directly to machines whereas the others can only be used in a comparatively indirect argument; for instance, if Godel’s theorem is to be used we need in addition to have some means of describing logical systems in terms of machines, and machines in terms of logical systems. The result in question refers to a type of machine which is essentially a digital computer with an infinite capacity. It states that there are certain things that such a machine cannot do. If it is rigged up to give answers to questions as in the imitation game, there will be some questions to which it will either give a wrong answer, or fail to give an answer at all, however much time is allowed for a reply. There may, of course, be many such questions, and questions which cannot be answered by one machine may be satisfactorily answered by another. We are of course supposing for the present that the questions are of the kind to which an answer “Yes” or “No” is appropriate, rather than questions such as “What do you think of Picasso?” The questions that we know the machines must fail on are of this type, “Consider the machine specified as follows…Will this machine ever answer ‘Yes’ to any question?” The dots are to be replaced by a description of some machine in a standard form, which could be something like that used in Section 6.5. When the machine described bears a certain comparatively simple relation to the machine which is under interrogation, it can be shown that the answer is either wrong or not forthcoming. This is the mathematical result; it is argued that it proves a disability of machines to which the human intellect is not subject.

对这一论点的简短回答是，尽管任何特定机器的能力都是有限的，但人们只是说，人类的智力不存在这种限制，而没有任何证据。但我认为不能如此轻易地否定这一观点。每当这些机器中的一台被问到适当的关键问题并给出明确的答案时，我们就知道这个答案一定是错的，这给了我们某种优越感。这种感觉是虚幻的吗？毫无疑问，这是非常真实的，但我认为不应该太过重视它。我们自己也经常对问题给出错误的答案，因此没有理由对机器的这种易错性感到高兴。此外，只有在这种场合，我们才能感受到我们的优越性，相对于我们取得小小胜利的那台机器。不可能同时战胜所有机器。简而言之，可能有人比任何给定的机器更聪明，但也可能还有其他机器更聪明，等等。

The short answer to this argument is that, although it is established that there are limitations to the powers of any particular machine, it has only been stated, without any sort of proof, that no such limitations apply to the human intellect. But I do not think this view can be dismissed quite so lightly. Whenever one of these machines is asked the appropriate critical question, and gives a definite answer, we know that this answer must be wrong, and this gives us a certain feeling of superiority. Is this feeling illusory? It is no doubt quite genuine, but I do not think too much importance should be attached to it. We too often give wrong answers to questions ourselves to be justified in being very pleased at such evidence of fallibility on the part of the machines. Further, our superiority can only be felt on such an occasion in relation to the one machine over which we have scored our petty triumph. There would be no question of triumphing simultaneously over all machines. In short, then, there might be men cleverer than any given machine, but then again there might be other machines cleverer again, and so on.

我认为，那些坚持数学论证的人大多愿意接受模仿游戏作为讨论的基础。那些相信前两种反对意见的人可能对任何标准都不感兴趣。

Those who hold to the mathematical argument would, I think, mostly be willing to accept the imitation game as a basis for discussion. Those who believe in the two previous objections would probably not be interested in any criteria.

（4）从意识出发的论证。这一论证在杰斐逊教授 1949 年的李斯特演说中得到了很好的表达，我从中引用了这段话。

(4) THE ARGUMENT FROM CONSCIOUSNESS. This argument is very well expressed in Professor Jefferson’s Lister Oration for 1949, from which I quote.

这种论证似乎否定了我们测试的有效性。根据这种观点的最极端形式，人们唯一能确定机器会思考的方式就是成为机器并感受到自己在思考。然后人们可以向世界描述这些感觉，但当然没有人有理由去注意。同样，根据这种观点，知道一个人会思考的唯一方式就是成为那个人。这实际上是唯我论的观点。这可能是最合乎逻辑的观点，但它使思想交流变得困难。A 倾向于相信“A 思考但 B 不思考”，而 B 则相信“B 思考但 A 不思考”。与其就这一点不断争论，不如通常采用每个人都思考的礼貌惯例。

This argument appears to be a denial of the validity of our test. According to the most extreme form of this view, the only way by which one could be sure that a machine thinks is to be the machine and to feel oneself thinking. One could then describe these feelings to the world, but of course no one would be justified in taking any notice. Likewise according to this view, the only way to know that a man thinks is to be that particular man. It is in fact the solipsist point of view. It may be the most logical view to hold but it makes communication of ideas difficult. A is liable to believe “A thinks but B does not” while B believes “B thinks but A does not”. Instead of arguing continually over this point, it is usual to have the polite convention that everyone thinks.

我确信杰斐逊教授不想采取极端和唯我论的观点。他可能很愿意接受模仿游戏作为一种测试。这种游戏（省略了玩家 B）在实践中经常以口头答辩的名义使用，以发现某人是否真正理解了某件事，还是“鹦鹉学舌”。让我们听听这种口头答辩的一部分：

I am sure that Professor Jefferson does not wish to adopt the extreme and solipsist point of view. Probably he would be quite willing to accept the imitation game as a test. The game (with the player B omitted) is frequently used in practice under the name of viva voce to discover whether someone really understands something or has “learned it parrot fashion”. Let us listen in to a part of such a viva voce:

等等。如果十四行诗写作机器能够在口试中像这样回答，杰斐逊教授会怎么说？我不知道他是否会认为机器“只是人工发出信号”这些答案，但如果答案像上文一样令人满意和持久，我不认为他会把它描述为“一个简单的发明”。我认为，这句话意在涵盖​​这样的设备，例如在机器中包含某人阅读十四行诗的记录，并适当切换以不时打开它。

And so on. What would Professor Jefferson say if the sonnet-writing machine were able to answer like this in the viva voce? I do not know whether he would regard the machine as “merely artificially signaling” these answers, but if the answers were as satisfactory and sustained as in the above passage I do not think he would describe it as “an easy contrivance”. This phrase is, I think, intended to cover such devices as the inclusion in the machine of a record of someone reading a sonnet, with appropriate switching to turn it on from time to time.

简而言之，我认为大多数支持意识论证的人可以被说服放弃它，而不是被迫接受唯我论立场。然后他们可能愿意接受我们的测试。

In short then, I think that most of those who support the argument from consciousness could be persuaded to abandon it rather than be forced into the solipsist position. They will then probably be willing to accept our test.

我不想给人留下这样的印象：我认为意识没有任何神秘之处。例如，任何试图确定意识的尝试都会带来某种悖论。但我认为，在我们回答本文所关注的问题之前，不一定需要解决这些谜团。

I do not wish to give the impression that I think there is no mystery about consciousness. There is, for instance, something of a paradox connected with any attempt to localize it. But I do not think these mysteries necessarily need to be solved before we can answer the question with which we are concerned in this paper.

（5）各种缺陷的论据。这些论据的形式是：“我承认你可以让机器做你提到的所有事情，但你永远无法让机器做 X。”在这方面提出了许多 X 特征。我提供一些选择：

(5) ARGUMENTS FROM VARIOUS DISABILITIES. These arguments take the form, “I grant you that you can make machines do all the things you have mentioned but you will never be able to make one to do X.” Numerous features X are suggested in this connection. I offer a selection:

通常，这些说法没有得到任何支持。我相信它们大多建立在科学归纳原理的基础上。一个人一生中见过成千上万台机器。从他所看到的机器中，他得出了许多一般性结论。它们很丑陋，每台机器的设计目的都非常有限，当需要用于细微不同的目的时，它们毫无用处，其中任何一台机器的行为变化都非常小，等等。自然，他得出结论，这些是机器的必要属性。这些限制中的许多都与大多数机器的存储容量非常小有关。（我假设存储容量的概念以某种方式扩展到涵盖离散状态机以外的机器。确切的定义并不重要，因为目前的讨论并不要求数学准确性。）几年前，当人们对数字计算机知之甚少时，如果有人提到它们的特性而不描述它们的构造，就有可能引起人们对它们的怀疑。这大概是由于科学归纳原理的类似应用。这些原理的应用当然在很大程度上是无意识的。当一个被烧伤的孩子害怕火，并通过躲避火来表明他害怕火时，我应该说他在运用科学归纳法。（当然，我也可以用许多其他方式来描述他的行为。）人类的工作和习俗似乎不是应用科学归纳法的合适材料。如果要获得可靠的结果，必须调查很大一部分时空。否则，我们可能会（像大多数英国孩子一样）认定每个人都说英语，学习法语是愚蠢的。

No support is usually offered for these statements. I believe they are mostly founded on the principle of scientific induction. A man has seen thousands of machines in his lifetime. From what he sees of them he draws a number of general conclusions. They are ugly, each is designed for a very limited purpose, when required for a minutely different purpose they are useless, the variety of behavior of any one of them is very small, and so on and so forth. Naturally he concludes that these are necessary properties of machines in general. Many of these limitations are associated with the very small storage capacity of most machines. (I am assuming that the idea of storage capacity is extended in some way to cover machines other than discrete state machines. The exact definition does not matter as no mathematical accuracy is claimed in the present discussion.) A few years ago, when very little had been heard of digital computers, it was possible to elicit much incredulity concerning them, if one mentioned their properties without describing their construction. That was presumably due to a similar application of the principle of scientific induction. These applications of the principle are of course largely unconscious. When a burned child fears the fire and shows that he fears it by avoiding it, I should say that he was applying scientific induction. (I could of course also describe his behavior in many other ways.) The works and customs of mankind do not seem to be very suitable material to which to apply scientific induction. A very large part of space-time must be investigated if reliable results are to be obtained. Otherwise we may (as most English children do) decide that everybody speaks English, and that it is silly to learn French.

然而，对于已提到的许多缺陷，需要特别指出。无法享用草莓和奶油可能让读者觉得是轻浮的。也许可以制造一台机器来享用这道美味佳肴，但任何让机器这样做的尝试都是愚蠢的。这种缺陷的重要之处在于，它会导致其他一些缺陷，例如，导致人与机器之间难以像白人与白人之间或黑人与黑人之间那样产生友好关系。

There are, however, special remarks to be made about many of the disabilities that have been mentioned. The inability to enjoy strawberries and cream may have struck the reader as frivolous. Possibly a machine might be made to enjoy this delicious dish, but any attempt to make one do so would be idiotic. What is important about this disability is that it contributes to some of the other disabilities, for instance, to the difficulty of the same kind of friendliness occurring between man and machine as between white man and white man, or between black man and black man.

“机器不会犯错”的说法似乎很奇怪。人们忍不住要反驳：“难道它们就因此更糟糕吗？”但让我们采取一种更同情的态度，试着看看这到底是什么意思。我认为这种批评可以用模仿游戏来解释。据称，询问者只需给机器出一些算术题就能区分机器和人。机器会因为其致命的准确性而被揭穿。对此的回答很简单。机器（为玩游戏而编程）不会试图给出算术问题的正确答案。它会故意以一种精心设计的方式引入错误，以迷惑询问者。机械故障可能会通过对算术中应该犯什么样的错误的不适当决定而表现出来。即使是这种解释的批评并不够同情。但我们没有足够的篇幅来进一步探讨它。在我看来，这种批评取决于两种错误的混淆。我们可以称它们为“功能错误”和“结论错误”。功能错误是由于某些机械或电气故障导致机器的行为与设计时不同。在哲学讨论中，人们喜欢忽略这种错误的可能性；因此人们讨论的是“抽象机器”。这些抽象机器是数学虚构，而不是物理对象。根据定义，它们不会出现功能错误。从这个意义上说，我们可以真正地说“机器永远不会犯错误”。结论错误只有在机器的输出信号具有某种含义时才会出现。例如，机器可能会打出数学方程式或英文句子。当输入错误的命题时，我们说机器犯了结论错误。显然，没有任何理由说机器不会犯这种错误。它可能什么也不做，只是反复输入“0 = 1”。举一个不那么反常的例子，它可能有一些通过科学归纳得出结论的方法。我们必须预料到，这种方法偶尔会导致错误的结果。

The claim that “machines cannot make mistakes” seems a curious one. One is tempted to retort, “Are they any the worse for that?” But let us adopt a more sympathetic attitude, and try to see what is really meant. I think this criticism can be explained in terms of the imitation game. It is claimed that the interrogator could distinguish the machine from the man simply by setting them a number of problems in arithmetic. The machine would be unmasked because of its deadly accuracy. The reply to this is simple. The machine (programmed for playing the game) would not attempt to give the right answers to the arithmetic problems. It would deliberately introduce mistakes in a manner calculated to confuse the interrogator. A mechanical fault would probably show itself through an unsuitable decision as to what sort of a mistake to make in the arithmetic. Even this interpretation of the criticism is not sufficiently sympathetic. But we cannot afford the space to go into it much further. It seems to me that this criticism depends on a confusion between two kinds of mistakes. We may call them “errors of functioning” and “errors of conclusion”. Errors of functioning are due to some mechanical or electrical fault which causes the machine to behave otherwise than it was designed to do. In philosophical discussions one like to ignore the possibility of such errors; one is therefore discussing “abstract machines”. These abstract machines are mathematical fictions rather than physical objects. By definition they are incapable of errors of functioning. In this sense we can truly say that “machines can never make mistakes”. Errors of conclusion can only arise when some meaning is attached to the output signals from the machine. The machine might, for instance, type out mathematical equations, or sentences in English. When a false proposition is typed we say that the machine has committed an error of conclusion. There is clearly no reason at all for saying that a machine cannot make this kind of mistake. It might do nothing but type out repeatedly “0 = 1”. To take a less perverse example, it might have some method for drawing conclusions by scientific induction. We must expect such a method to lead occasionally to erroneous results.

当然，只有当可以证明机器对某些主题有某种思考时，才能回答“机器不能成为其自身思想的主体”这一说法。然而，“机器操作的主题”似乎确实意味着什么，至少对处理它的人来说是这样。例如，如果机器试图找到方程x ² − 40 x − 11 = 0 的解，人们就会倾向于将这个方程描述为当时机器主题的一部分。它可能被用来帮助编写自己的程序，或者预测自身结构改变的影响。通过观察自身行为的结果，它可以修改自己的程序，以更有效地实现某些目的。这些都是不久的将来的可能性，而不是乌托邦式的梦想。

The claim that a machine cannot be the subject of its own thought can of course only be answered if it can be shown that the machine has some thought with some subject matter. Nevertheless, “the subject matter of a machine’s operations” does seem to mean something, at least to the people who deal with it. If, for instance, the machine were trying to find a solution of the equation x² − 40x− 11 = 0, one would be tempted to describe this equation as part of the machine’s subject matter at that moment. It may be used to help in making up its own programs, or to predict the effect of alterations in its own structure. By observing the results of its own behavior it can modify its own programs so as to achieve some purpose more effectively. These are possibilities of the near future, rather than Utopian dreams.

有人批评机器不能有太多的行为多样性，其实就是说机器不能有太大的存储容量。直到最近，存储容量达到一千位数字的机器都十分罕见。

The criticism that a machine cannot have much diversity of behavior is just a way of saying that it cannot have much storage capacity. Until fairly recently a storage capacity of even a thousand digits was very rare.

我们在这里考虑的批评往往是意识论证的伪装形式。通常，如果有人坚持认为机器可以做这些事情之一，并描述了机器可以使用的方法，那么他不会给人留下太深刻的印象。人们认为这种方法（无论它是什么，因为它必须是机械的）实际上相当低劣。比较一下上面引用的杰斐逊声明中的括号。

The criticisms that we are considering here are often disguised forms of the argument from consciousness. Usually if one maintains that a machine can do one of these things, and describes the kind of method that the machine could use, one will not make much of an impression. It is thought that the method (whatever it may be, for it must be mechanical) is really rather base. Compare the parenthesis in Jefferson’s statement quoted above.

（6）洛夫莱斯夫人的反对意见。关于巴贝奇分析机的最详细信息来自洛夫莱斯夫人的回忆录（1842 年）。在回忆录中，她指出：“分析机不打算创造任何东西。只要我们知道命令它做什么，它就能做什么”（斜体字为她所加）。哈特里（1949 年）引用了这句话，并补充道：“这并不意味着不可能制造出‘为自己思考’的电子设备，或者从生物学角度来说，人们可以建立条件反射，作为‘学习’的基础。这在原则上是否可行是一个令人振奋和激动人心的问题，最近的一些发展提出了这个问题。但当时建造或设计的机器似乎不具备这种特性。”

(6) LADY LOVELACE’S OBJECTION. Our most detailed information of Babbage’s Analytical Engine comes from a memoir by Lady Lovelace (1842). In it she states, “The Analytical Engine has no pretensions to originate anything. It can do whatever we know how to order it to perform” (her italics). This statement is quoted by Hartree (1949) who adds: “This does not imply that it may not be possible to construct electronic equipment which will ‘think for itself’, or in which, in biological terms, one could set up a conditioned reflex, which would serve as a basis for ‘learning’. Whether this is possible in principle or not is a stimulating and exciting question, suggested by some of these recent developments. But it did not seem that the machines constructed or projected at the time had this property.”

我完全同意 Hartree 的观点。值得注意的是，他并没有断言所讨论的机器不具备这种属性，而是洛夫莱斯夫人掌握的证据并没有鼓励她相信它们具有这种属性。很有可能，所讨论的机器在某种意义上具有这种属性。假设某个离散状态机具有这种属性。分析机是一种通用数字计算机，因此，如果其存储容量和速度足够，它可以通过适当的编程来模仿所讨论的机器。伯爵夫人或巴贝奇可能没有想到这个论点。无论如何，他们没有义务声称所有可以声称的东西。

I am in thorough agreement with Hartree over this. It will be noticed that he does not assert that the machines in question had not got the property, but rather that the evidence available to Lady Lovelace did not encourage her to believe that they had it. It is quite possible that the machines in question had in a sense got this property. For suppose that some discrete state machine has the property. The Analytical Engine was a universal digital computer, so that, if its storage capacity and speed were adequate, it could by suitable programming be made to mimic the machine in question. Probably this argument did not occur to the Countess or to Babbage. In any case there was no obligation on them to claim all that could be claimed.

整个问题将在学习机器的标题下再次被考虑。

This whole question will be considered again under the heading of learning machines.

洛夫莱斯夫人的反对意见的一个变体是，机器“永远不能做任何真正新颖的事情”。这句话可以暂时用“太阳底下没有新鲜事”来回避。谁能确定他所做的“原创工作”不是通过教学在他身上种下的种子的生长，也不是遵循众所周知的一般原则的结果。反对意见的一个更好的变体是，机器永远不会“让我们措手不及”。这句话是一个更直接的挑战，可以直接应对。机器经常让我措手不及。这主要是因为我没有做足够的计算来决定期望它们做什么，或者更确切地说，因为虽然我做了计算，但我做得太匆忙、马虎了，冒了风险。也许我对自己说，“我想这里的电压应该和那里的电压一样：无论如何，让我们假设它是一样。”当然，我经常是错的，结果让我感到惊讶，因为到实验结束时，这些假设已经被遗忘了。这些承认使我容易受到有关我恶毒行径的说教，但当我证明我所经历的意外时，我的可信度却毫不怀疑。

A variant of Lady Lovelace’s objection states that a machine can “never do anything really new”. This may be parried for moment with the saw, “There is nothing new under the sun.” Who can be certain that “original work” that he has done was not simply the growth of the seed planted in him by teaching, or the effect of following well-known general principles. A better variant of the objection says that a machine can never “take us by surprise”. This statement is a more direct challenge and can be met directly. Machines take me by surprise with great frequency. This is largely because I do not do sufficient calculation to decide what to expect them to do, or rather because, although I do a calculation, I do it in a hurried, slipshod fashion, taking risks. Perhaps I say to myself, “I suppose the voltage here ought to be the same as there: anyway let’s assume it is.” Naturally I am often wrong, and the result is a surprise for me, for by the time the experiment is done these assumptions have been forgotten. These admissions lay me open to lectures on the subject of my vicious ways, but do not throw any doubt on my credibility when I testify to the surprises I experience.

我不指望这个答复能让批评者闭嘴。他可能会说，这种惊奇是由于我某种创造性的心理活动，并不代表机器的功劳。这又把我们带回到意识论证，远离惊奇的概念。我们必须认为这条论证已经结束，但也许值得一提的是，欣赏某种令人惊讶的东西需要同样多的“创造性心理活动”，无论令人惊讶的事件是来自人、书、机器还是其他任何东西。

I do not expect this reply to silence my critic. He will probably say that such surprises are due to some creative mental act on my part, and reflect no credit on the machine. This leads us back to the argument from consciousness, and far from the idea of surprise. It is a line of argument we must consider closed, but it is perhaps worth remarking that the appreciation of something as surprising requires as much of a “creative mental act” whether the surprising event originates from a man, a book, a machine or anything else.

我认为，认为机器无法制造意外的观点是哲学家和数学家特别容易犯的一个错误。这种假设认为，只要一个事实出现在大脑中，该事实的所有后果都会同时出现在大脑中。在许多情况下，这是一个非常有用的假设，但人们很容易忘记它是错误的。这样做的自然结果是，人们认为仅仅从数据和一般原则中得出后果是毫无意义的。

The view that machines cannot give rise to surprises is due, I believe, to a fallacy to which philosophers and mathematicians are particularly subject. This is the assumption that as soon as a fact is presented to a mind all consequences of that fact spring into the mind simultaneously with it. It is a very useful assumption under many circumstances, but one too easily forgets that it is false. A natural consequence of doing so is that one then assumes that there is no virtue in the mere working out of consequences from data and general principles.

（7）神经系统连续性的论证。神经系统当然不是离散状态机。神经脉冲大小信息中的一个小错误，可能会对输出脉冲的大小产生很大的影响。有人可能会说，既然如此，就不能指望用离散状态系统来模仿神经系统的行为。

(7) ARGUMENT FROM CONTINUITY IN THE NERVOUS SYSTEM. The nervous system is certainly not a discrete state machine. A small error in the information about the size of a nervous impulse impinging on a neuron, may make a large difference to the size of the outgoing impulse. It may be argued that, this being so, one cannot expect to be able to mimic the behavior of the nervous system with a discrete state system.

确实，离散状态机必须不同于连续状态机。但是，如果我们遵守模仿游戏的条件，询问者将无法利用这种差异。如果我们考虑其他更简单的连续机器，情况会更清楚。微分分析仪就很好用。（微分分析仪是一种机器，不是离散状态类型，用于某些类型的计算。）其中一些以打字形式提供答案，因此适合参与游戏。数字计算机不可能准确预测微分分析仪会对问题给出什么答案，但它完全有能力给出正确的答案。例如，如果要求给出π的值（实际上约为3.1416），则在3.12、3.13、3.14、3.15、3.16之间随机选择是合理的，概率为0.05、0.15、0.55、0.19、0.06（假设）。在这种情况下，询问者很难区分微分分析仪和数字计算机。

It is true that a discrete state machine must be different from a continuous machine. But if we adhere to the conditions of the imitation game, the interrogator will not be able to take any advantage of this difference. The situation can be made clearer if we consider some other simpler continuous machine. A differential analyzer will do very well. (A differential analyzer is a certain kind of machine, not of the discrete state type, used for some types of calculation.) Some of these provide their answers in a typed form, and so are suitable for taking part in the game. It would not be possible for a digital computer to predict exactly what answers the differential analyzer would give to a problem, but it would be quite capable of giving the right sort of answer. For instance, if asked to give the value of π (actually about 3.1416) it would be reasonable to choose at random between the values 3.12, 3.13, 3.14, 3.15, 3.16 with the probabilities of 0.05, 0.15, 0.55, 0.19, 0.06 (say). Under these circumstances it would be very difficult for the interrogator to distinguish the differential analyzer from the digital computer.

（8）行为非形式性论证。不可能制定一套规则来描述一个人在每一种可以想象到的情况下应该做什么。例如，人们可以制定一条规则，即看到红灯时要停车，看到绿灯时要通行；但如果由于某种错误，两种灯同时出现怎么办？人们也许会决定停车是最安全的。但这个决定以后可能会引发进一步的困难。试图提供涵盖所有可能发生的情况的行为规则，甚至包括由交通信号灯引起的情况，似乎是不可能的。我同意所有这些。

(8) THE ARGUMENT FROM INFORMALITY OF BEHAVIOR. It is not possible to produce a set of rules purporting to describe what a man should do in every conceivable set of circumstances. One might for instance have a rule that one is to stop when one sees a red traffic light, and to go if one sees a green one; but what if by some fault both appear together? One may perhaps decide that it is safest to stop. But some further difficulty may well arise from this decision later. To attempt to provide rules of conduct to cover every eventuality, even those arising from traffic lights, appears to be impossible. With all this I agree.

由此，有人认为我们不能成为机器。我将尝试重现这一论点，但恐怕很难公正地阐述它。它似乎是这样的：“如果每个人都有一套明确的行为准则来规范自己的生活，那么他就和机器没什么两样。但是没有这样的规则，所以人不能成为机器。”中间的未分配部分很明显。我认为这个论点从来没有像这样提出过，但我相信这是使用的论点。然而，“行为规则”和“行为法则”之间可能存在一定的混淆，使问题变得模糊。我所说的“行为规则”是指诸如“看到红灯就停车”之类的戒律，人们可以据此采取行动，并且可以意识到这一点。我所说的“行为法则”是指应用于人体的自然法则，例如“如果你掐他，他会吱吱叫”。如果我们将引用的论证中的“规范其生活的行为法则”替换为“规范其生活的行为法则”，未分配的中间部分就不再是不可克服的。因为我们相信，不仅受行为法则的约束意味着成为某种机器（尽管不一定是离散状态机），而且反过来，成为这样的机器意味着受此类法则的约束。然而，我们无法像相信没有完整的行为规则那样轻易地说服自己没有完整的行为法则。我们知道找到此类法则的唯一方法是科学观察，而且我们当然不知道在什么情况下我们可以说：“我们已经搜索得足够多了。没有这样的法则。”

From this it is argued that we cannot be machines. I shall try to reproduce the argument, but I fear I shall hardly do it justice. It seems to run something like this: “If each man had a definite set of rules of conduct by which he regulated his life he would be no better than a machine. But there are no such rules, so men cannot be machines.” The undistributed middle is glaring. I do not think the argument is ever put quite like this, but I believe this is the argument used nevertheless. There may however be a certain confusion between “rules of conduct” and “laws of behavior” to cloud the issue. By “rules of conduct” I mean precepts such as “Stop if you see red lights”, on which one can act, and of which one can be conscious. By “laws of behavior” I mean laws of nature as applied to a man’s body such as “if you pinch him he will squeak”. If we substitute “laws of behavior which regulate his life” for “laws of conduct by which he regulates his life” in the argument quoted the undistributed middle is no longer insuperable. For we believe that it is not only true that being regulated by laws of behavior implies being some sort of machine (though not necessarily a discrete state machine), but that conversely being such a machine implies being regulated by such laws. However, we cannot so easily convince ourselves of the absence of complete laws of behavior as of complete rules of conduct. The only way we know of for finding such laws is scientific observation, and we certainly know of no circumstances under which we could say: “We have searched enough. There are no such laws.”

我们可以更有力地证明，任何此类说法都是不合理的。假设我们肯定能找到这样的定律（如果它们存在的话）。那么，给定一个离散状态机，通过观察它，我们肯定能够发现足够的信息来预测它的未来行为，而且是在合理的时间内，比如说一千年。但事实似乎并非如此。我在曼彻斯特计算机上设置了一个只使用 1000 个存储单元的小程序，机器在两秒钟内用一个十六位数字回复另一个十六位数字。我敢说没有人能从这些回复中学到足够的信息来预测对未尝试过的值的任何回复。

We can demonstrate more forcibly that any such statement would be unjustified. For suppose we could be sure of finding such laws if they existed. Then given a discrete state machine it should certainly be possible to discover by observation sufficient about it to predict its future behavior, and this within a reasonable time, say a thousand years. But this does not seem to be the case. I have set up on the Manchester computer a small program using only 1000 units of storage, whereby the machine supplied with one sixteen figure number replies with another within two seconds. I would defy anyone to learn from these replies sufficient about the program to be able to predict any replies to untried values.

（9）超感官知觉论证。我假设读者熟悉超感官知觉的概念，以及它的四个项目的含义，即心灵感应、透视、预知和意志力。这些令人不安的现象似乎否定了我们所有的常规科学思想。我们多么想诋毁它们！不幸的是，至少对于心灵感应来说，统计证据是压倒性的。很难重新整理一个人的想法以适应这些新事实。一旦接受了它们，相信鬼魂和妖怪似乎并不是一个很大的步骤。我们的身体只是按照已知的物理定律以及其他一些尚未发现但有些相似的定律运动的想法将是最先被抛弃的想法之一。

(9) THE ARGUMENT FROM EXTRA-SENSORY PERCEPTION. I assume that the reader is familiar with the idea of extra-sensory perception, and the meaning of the four items of it, namely, telepathy, clairvoyance, precognition and psychokinesis. These disturbing phenomena seem to deny all our usual scientific ideas. How we should like to discredit them! Unfortunately the statistical evidence, at least for telepathy, is overwhelming. It is very difficult to rearrange one’s ideas so as to fit these new facts in. Once one has accepted them it does not seem a very big step to believe in ghosts and bogies. The idea that our bodies move simply according to the known laws of physics, together with some others not yet discovered but somewhat similar, would be one of the first to go.

在我看来，这个论点相当有力。人们可以反驳说，许多科学理论似乎在实践中仍然可行，尽管与 ESP 相冲突；事实上，如果人们忘记了它，人们可以过得很好。这相当令人安慰，人们担心思考正是 ESP 可能特别相关的那种现象。

This argument is to my mind quite a strong one. One can say in reply that many scientific theories seem to remain workable in practice, in spite of clashing with E.S.P.; that in fact one can get along very nicely if one forgets about it. This is rather cold comfort, and one fears that thinking is just the kind of phenomenon where E.S.P. may be especially relevant.

基于 ESP 的更具体论点可能是这样的：“让我们玩模仿游戏，使用一个擅长心灵感应接收的人和一台数字计算机作为证人。询问者可以问这样的问题：‘我右手的牌属于什么花色？’这个人通过心灵感应或千里眼在 400 张牌中给出了 130 次正确答案。机器只能随机猜测，可能猜对 104 次，因此询问者做出了正确的识别。” 这里有一个有趣的可能性。假设数字计算机包含一个随机数生成器。那么使用它来决定给出什么答案将是自然而然的。但随机数生成器将受制于询问者的心灵感应能力。也许这种心灵感应可能会导致机器比概率计算预期的更频繁地猜对，因此询问者可能仍然无法做出正确的识别。另一方面，他可能能够通过千里眼在没有任何疑问的情况下猜对。有了 ESP，一切事情都可能发生。

A more specific argument based on E.S.P. might run as follows: “Let us play the imitation game, using as witnesses a man who is good as a telepathic receiver, and a digital computer. The interrogator can ask such questions as ‘What suit does the card in my right hand belong to?’The man by telepathy or clairvoyance gives the right answer 130 times out of 400 cards. The machine can only guess at random, and perhaps get 104 right, so the interrogator makes the right identification.” There is an interesting possibility which opens here. Suppose the digital computer contains a random number generator. Then it will be natural to use this to decide what answer to give. But then the random number generator will be subject to the psychokinetic powers of the interrogator. Perhaps this psychokinesis might cause the machine to guess right more often then would be expected on a probability calculation, so that the interrogator might still be unable to make the right identification. On the other hand, he might be able to guess right without any questioning, by clairvoyance. With E.S.P. anything may happen.

如果心灵感应被认可，就需要加强我们的测试。这种情况可以被视为类似于审讯者自言自语，而其中一名参赛者正把耳朵贴在墙上听的情况。将参赛者放入“防心灵感应室”将满足所有要求。

If telepathy is admitted it will be necessary to tighten our test. The situation could be regarded as analogous to that which would occur if the interrogator were talking to himself and one of the competitors was listening with his ear to the wall. To put the competitors into a “telepathy-proof room” would satisfy all requirements.

6.7 学习机器

6.7  Learning machines

读者应该已经预料到，我没有非常令人信服的正面论据来支持我的观点。如果我有，我就不会费尽心思去指出相反观点的谬误。我现在将给出我所掌握的证据。

The reader will have anticipated that I have no very convincing arguments of a positive nature to support my views. If I had I should not have taken such pains to point out the fallacies in contrary views. Such evidence as I have I shall now give.

让我们回到洛夫莱斯夫人的反对意见，她指出机器只能按照我们的指令行事。有人可能会说，一个人可以“注入”一个想法到机器中，机器会做出一定程度的反应，然后进入静止状态，就像被锤子敲击的钢琴弦一样。另一个比喻是小于临界尺寸的原子堆：注入的想法相当于从外部进入堆中的中子。每个这样的中子都会造成一定的扰动，最终会消失。然而，如果堆的大小足够大，这种进入中子引起的扰动很可能会持续下去，直到整个堆被摧毁。心灵有相应的现象吗？机器有相应的现象吗？人类心灵似乎确实有相应的现象。它们中的大多数似乎是“亚临界的”，也就是说，在这个类比中对应于亚临界尺寸的堆。向这样的心灵提出一个想法，平均而言，它会产生不到一个想法作为回应。一小部分是超临界的。向这样的头脑提出一个想法可能会产生一整套“理论”，其中包括二级、三级和更远的思想。动物的头脑似乎非常肯定是亚临界的。坚持这个类比，我们问，“机器能被制造成超临界的吗？”

Let us return for a moment to Lady Lovelace’s objection, which stated that the machine can only do what we tell it to do. One could say that a man can “inject” an idea into the machine, and that it will respond to a certain extent and then drop into quiescence, like a piano string struck by a hammer. Another simile would be an atomic pile of less than critical size: an injected idea is to correspond to a neutron entering the pile from without. Each such neutron will cause a certain disturbance which eventually dies away. If, however, the size of the pile is sufficiently increased, the disturbance caused by such an incoming neutron will very likely go on and on, increasing until the whole pile is destroyed. Is there a corresponding phenomenon for minds, and is there one for machines? There does seem to be one for the human mind. The majority of them seem to be “subcritical”, that is, to correspond in this analogy to piles of subcritical size. An idea presented to such a mind will on an average give rise to less than one idea in reply. A smallish proportion are supercritical. An idea presented to such a mind may give rise to a whole “theory” consisting of secondary, tertiary and more remote ideas. Animals’ minds seem to be very definitely subcritical. Adhering to this analogy we ask, “Can a machine be made to be supercritical?”

“洋葱皮”的类比也很有帮助。在考虑心灵或大脑的功能时，我们发现某些操作可以用纯机械术语来解释。我们说这并不对应于真正的心灵：这是一种皮肤，如果我们要找到真正的心灵，就必须剥去它。但是，在剩下的部分中，我们发现了另一层需要剥去的皮，等等。以这种方式继续下去，我们是否会得到“真正的”心灵，还是最终会得到里面什么都没有的皮肤？在后一种情况下，整个心灵都是机械的。（然而，它不会是一台离散状态机。我们已经讨论过这个问题。）

The “skin of an onion” analogy is also helpful. In considering the functions of the mind or the brain we find certain operations which we can explain in purely mechanical terms. This we say does not correspond to the real mind: it is a sort of skin which we must strip off if we are to find the real mind. But then in what remains we find a further skin to be stripped off, and so on. Proceeding in this way, do we ever come to the “real” mind, or do we eventually come to the skin which has nothing in it? In the latter case the whole mind is mechanical. (It would not be a discrete state machine however. We have discussed this.)

最后两段话并不算是令人信服的论据。它们应该被描述为“倾向于产生信念的背诵”。

These last two paragraphs do not claim to be convincing arguments. They should rather be described as “recitations tending to produce belief”.

对于第 6.6 节开头表达的观点，唯一真正令人满意的支持是等待本世纪末，然后进行所描述的实验。但与此同时我们能说什么呢？如果实验要成功，现在应该采取什么步骤？

The only really satisfactory support that can be given for the view expressed at the beginning of Section 6.6 will be that provided by waiting for the end of the century and then doing the experiment described. But what can we say in the meantime? What steps should be taken now if the experiment is to be successful?

正如我所解释的，问题主要在于编程。工程学也必须取得进步，但这些进步似乎不足以满足要求。大脑存储容量的估计值从 10 ¹⁰到 10 ^{15 个}二进制数字不等。我倾向于较低的值，并相信只有很小一部分用于高级思维。大部分可能用于保留视觉印象。如果令人满意地玩模仿游戏需要超过 10 ^{9 个，我会感到惊讶，至少对一个盲人而言是这样。（注：《}大英百科全书》第 11 版的容量为 2 × 10 ^{9 。）即使采用目前的技术，10 7}的存储容量也是非常可行的。可能根本不需要提高机器的运行速度。现代机器中可以被视为神经细胞类似物的部件的工作速度比后者快一千倍。这应该能提供一个“安全边际”，以弥补以多种方式引起的速度损失。那么我们的问题就是找出如何对这些机器进行编程来玩这个游戏。以我目前的工作速度，我每天可以编写大约一千个程序，因此，如果没有任何东西被扔进废纸篓，大约六十名工人在五十年内稳定工作就可以完成这项工作。似乎需要一些更快捷的方法。

As I have explained, the problem is mainly one of programming. Advances in engineering will have to made too, but it seems unlikely that these will not be adequate for the requirements. Estimates of the storage capacity of the brain vary from 10¹⁰ to 10¹⁵ binary digits. I incline to the lower values and believe that only a very small fraction is used for the higher types of thinking. Most of it is probably used for the retention of visual impressions. I should be surprised if more than 10⁹ was required for satisfactory playing of the imitation game, at any rate against a blind man. (Note: The capacity of the Encyclopedia Britannica, eleventh edition, is 2 × 10⁹.) A storage capacity of 10⁷ would be a very practicable possibility even by present techniques. It is probably not necessary to increase the speed of operations of the machines at all. Parts of modern machines which can be regarded as analogues of nerve cells work about a thousand times faster than the latter. This should provide a “margin of safety” which could cover losses of speed arising in many ways. Our problem then is to find out how to program these machines to play the game. At my present rate of working I produce about a thousand digits of program a day, so that about sixty workers, working steadily through the fifty years might accomplish the job, if nothing went into the wastepaper basket. Some more expeditious method seems desirable.

在试图模仿成年人心智的过程中，我们必然会认真思考导致其处于当前状态的过程。我们可能会注意到三个组成部分：

In the process of trying to imitate an adult human mind we are bound to think a good deal about the process which has brought it to the state that it is in. We may notice three components:

与其尝试编写一个程序来模拟成人思维，为什么不尝试编写一个程序来模拟儿童的思维呢？如果对这个程序进行适当的教育，人们就会获得成人的大脑。据推测，儿童的大脑就像从文具店买来的笔记本一样。几乎没有什么机械装置，还有很多空白页。（从我们的角度来看，机械装置和写作几乎是同义词。）我们希望儿童大脑中的机械装置很少，这样就可以很容易地对类似的东西进行编程。我们可以假设，作为初步估计，教育工作量与人类儿童的工作量大致相同。

Instead of trying to produce a program to simulate the adult mind, why not rather try to produce one which simulates the child’s? If this were then subjected to an appropriate course of education one would obtain the adult brain. Presumably the child-brain is something like a notebook as one buys it from the stationers. Rather little mechanism, and lots of blank sheets. (Mechanism and writing are from our point of view almost synonymous.) Our hope is that there is so little mechanism in the child-brain that something like it can be easily programmed. The amount of work in the education we can assume, as a first approximation, to be much the same as for the human child.

因此，我们将问题分为两个部分——儿童程序和教育过程。这两个部分仍然紧密相连。我们不能指望第一次尝试就能找到一个好的儿童机器。我们必须尝试教这样的机器，然后看看它学习得有多好。然后人们可以尝试另一种，看看它是更好还是更坏。这个过程和进化之间有着明显的联系，通过识别

We have thus divided our problem into two parts—the child-program and the education process. These two remain very closely connected. We cannot expect to find a good child-machine at the first attempt. One must experiment with teaching one such machine and see how well it learns. One can then try another and see if it is better or worse. There is an obvious connection between this process and evolution, by the identifications

然而，人们可能希望这个过程比进化更快。适者生存是一种衡量优势的缓慢方法。实验者通过运用智慧，应该能够加快这一速度。同样重要的是，他不受随机突变的限制。如果他能找出某些弱点的原因，他可能就能想出那种能改善它的突变。

One may hope, however, that this process will be more expeditious than evolution. The survival of the fittest is a slow method for measuring advantages. The experimenter, by the exercise of intelligence, should be able to speed it up. Equally important is the fact that he is not restricted to random mutations. If he can trace a cause for some weakness he can probably think of the kind of mutation which will improve it.

不可能将与对正常儿童完全相同的教学过程应用于机器。例如，它不会有腿，所以不能要求它出去装满煤斗。也许它没有眼睛。但无论这些缺陷如何通过巧妙的工程设计克服，都无法让这个生物上学而不让其他孩子过分取笑它。必须给它一些指导。我们不必太在意腿、眼睛等等。海伦·凯勒小姐的例子表明，只要师生之间可以通过某种方式进行双向交流，教育就可以进行。

It will not be possible to apply exactly the same teaching process to the machine as to a normal child. It will not, for instance, be provided with legs, so that it could not be asked to go out and fill the coal scuttle. Possibly it might not have eyes. But however well these deficiencies might be overcome by clever engineering, one could not send the creature to school without the other children making excessive fun of it. It must be given some tuition. We need not be too concerned about the legs, eyes, and so on. The example of Miss Helen Keller shows that education can take place provided that communication in both directions between teacher and pupil can take place by some means or other.

我们通常将惩罚和奖励与教学过程联系起来。一些简单的儿童机器可以根据这种原则构建或编程。机器必须如此构造，使得在惩罚信号出现之前不久发生的事件不太可能重复发生，而奖励信号则增加了导致惩罚信号的事件重复发生的概率。这些定义并不预设机器的任何感觉。我用一台这样的儿童机器做了一些实验，并成功地教会了它一些东西，但教学方法太不正统，实验不能被认为是真正成功的。

We normally associate punishments and rewards with the teaching process. Some simple child-machines can be constructed or programmed on this sort of principle. The machine has to be so constructed that events which shortly preceded the occurrence of a punishment-signal are unlikely to be repeated, whereas a reward-signal increases the probability of repetition of the events which led up to it. These definitions do not presuppose any feelings on the part of the machine. I have done some experiments with one such child-machine, and succeeded in teaching it a few things, but the teaching method was too unorthodox for the experiment to be considered really successful.

惩罚和奖励的使用最多只能成为教学过程的一部分。粗略地说，如果老师没有其他与学生交流的方式，那么他能获得的信息量不会超过所施加的奖励和惩罚的总数。当一个孩子学会背诵《卡萨布兰卡》时，如果只能通过“二十个问题”技术来发现文本，他可能会感到非常痛苦，每个“不”都以打击的形式出现。因此，有必要有其他一些“不带感情”的沟通渠道。如果有这些渠道，就可以通过惩罚和奖励来教机器服从用某种语言（例如符号语言）给出的命令。这些命令将通过“不带感情”的渠道传输。使用这种语言将大大减少所需的惩罚和奖励的数量。

The use of punishments and rewards can at best be a part of the teaching process. Roughly speaking, if the teacher has no other means of communicating to the pupil, the amount of information which can reach him does not exceed the total number of rewards and punishments applied. By the time a child has learned to repeat “Casablanca” he would probably feel very sore indeed, if the text could only be discovered by a “Twenty Questions” technique, every “No” taking the form of a blow. It is necessary therefore to have some other “unemotional” channels of communication. If these are available it is possible to teach a machine by punishments and rewards to obey orders given in some language, such as a symbolic language. These orders are to be transmitted through the “unemotional” channels. The use of this language will diminish greatly the number of punishments and rewards required.

关于儿童机器适合的复杂度，人们的看法可能有所不同。人们可能会尝试使其尽可能简单，并符合一般原则。或者，人们可能会“内置”一个完整的逻辑推理系统。2^在后一种情况下，存储库将主要包含定义和命题。命题将具有各种状态，例如已确定的事实、猜想、数学证明的定理、权威给出的陈述以及具有命题逻辑形式但没有信念值的表达式。某些命题可能被描述为“命令”。机器的构造应使得一旦命令被归类为“已确定”，就会自动执行相应的操作。为了说明这一点，假设老师对机器说：“现在做你的作业。”这可能会导致“老师说‘现在做你的作业’”被包括在已确定的事实中。另一个这样的事实可能是，“老师说的一切都是正确的”。将这些结合起来最终可能导致命令“现在做作业”被纳入既定事实，而根据机器的构造，这意味着作业实际上已经开始了；但效果并不令人满意。机器使用的推理过程不必满足最严格的逻辑学家的要求。例如，可能没有类型的层次结构。但这并不意味着会发生类型谬误，就像我们注定会从没有围栏的悬崖上掉下来一样。适当的命令（在系统内表达，不构成系统规则的一部分）例如“不要使用某个类，除非它是老师提到的类的子类”可以产生与“不要太靠近边缘”类似的效果。

Opinions may vary as to the complexity which is suitable in the child-machine. One might try to make it as simple as possible consistently with the general principles. Alternatively one might have a complete system of logical inference “built in”.² In the latter case the store would be largely occupied with definitions and propositions. The propositions would have various kinds of status, such as well-established facts, conjectures, mathematically proved theorems, statements given by an authority, and expressions having the logical form of a proposition but no belief-value. Certain propositions may be described as “imperatives”. The machine should be so constructed that as soon as an imperative is classed as “well-established” the appropriate action automatically takes place. To illustrate this, suppose the teacher says to the machine, “Do your homework now.” This may cause “Teacher says ‘Do your homework now’” to be included among the well-established facts. Another such fact might be, “Everything that teacher says is true”. Combining these may eventually lead to the imperative, “Do your homework now”, being included amongst the well-established facts, and this, by the construction of the machine, will mean that the homework actually gets started; but the effect is very unsatisfactory. The processes of inference used by the machine need not be such as would satisfy the most exacting logicians. There might, for instance, be no hierarchy of types. But this need not mean that type fallacies will occur, any more than we are bound to fall over unfenced cliffs. Suitable imperatives (expressed within the systems, not forming part of the rules of the system) such as “Do not use a class unless it is a subclass of one which has been mentioned by teacher” can have a similar effect to “Do not go too near the edge.”

没有四肢的机器能够遵循的命令必定具有相当的智力特征，如上面给出的例子（做作业）。这些命令中最重要的是那些规定相关逻辑系统规则应用顺序的命令。因为在使用逻辑系统的每个阶段，都存在大量可选步骤，就遵守逻辑系统规则而言，允许应用其中任何步骤。这些选择决定了聪明的推理者和愚蠢的推理者之间的区别，而不是健全的推理者和谬误的推理者之间的区别。导致这种命令的命题可能是“当提到苏格拉底时，使用芭芭拉的三段论”或“如果一种方法被证明比另一种方法更快，不要使用较慢的方法”。其中一些可能是“权威给出的”，但其他一些可能是由机器本身产生的，比如通过科学归纳法。

The imperatives that can be obeyed by a machine that has no limbs are bound to be of a rather intellectual character, as in the example (doing homework) given above. Important among such imperatives will be ones which regulate the order in which the rules of the logical system concerned are to be applied. For at each stage when one is using a logical system, there is a very large number of alternative steps, any of which one is permitted to apply, so far as obedience to the rules of the logical system is concerned. These choices make the difference between a brilliant and a footling reasoner, not the difference between a sound and a fallacious one. Propositions leading to imperatives of this kind might be “When Socrates is mentioned, use the syllogism in Barbara” or “If one method has been proved to be quicker than another, do not use the slower method.” Some of these may be “given by authority”, but others may be produced by the machine itself, say by scientific induction.

对于某些读者来说，学习机器的想法可能显得自相矛盾。机器的操作规则怎么会改变呢？它们应该完整地描述机器将如何反应，无论它的历史如何，无论它经历什么变化。因此，规则是相当不随时间变化的。这是完全正确的。对这一悖论的解释是，在学习过程中发生变化的规则是一种不那么自命不凡的规则，只声称具有短暂的有效性。读者可以将其与美国宪法进行比较。

The idea of a learning machine may appear paradoxical to some readers. How can the rules of operation of the machine change? They should describe completely how the machine will react whatever its history might be, whatever changes it might undergo. The rules are thus quite time-invariant. This is quite true. The explanation of the paradox is that the rules which get changed in the learning process are of a rather less pretentious kind, claiming only an ephemeral validity. The reader may draw a parallel with the Constitution of the United States.

学习型机器的一个重要特征是，它的老师通常对机器内部发生的事情一无所知，尽管他可能在某种程度上仍能预测学生的行为。这应该最适用于对经过精心设计（或程序）的儿童机器进行后期教育。这与使用机器进行计算时的正常程序形成鲜明对比：人们的目标是在计算的每个时刻对机器的状态有一个清晰的心理图像。这个目标只有经过努力才能实现。与此相比，“机器只能做我们知道如何命令它做的事情” ³的观点似乎很奇怪。我们可以放入机器中的大多数程序都会导致它做一些我们根本无法理解的事情，或者我们认为是完全随机的行为。智能行为可能与计算中涉及的完全有纪律的行为有所不同，但差异很小，不会引起随机行为或无意义的重复循环。通过教学和学习过程为我们的机器做好参与模仿游戏的准备，另一个重要结果是，“人类的易犯错误”很可能以一种相当自然的方式被承认，也就是说，无需特殊的“指导”。（读者应该将这一点与第 113 页的观点相协调。）学习的过程不会产生百分之百确定的结果；如果它们能产生百分之百确定的结果，那么它们就不可能被消除。

An important feature of a learning machine is that its teacher will often be very largely ignorant of quite what is going on inside, although he may still be able to some extent to predict his pupil’s behavior. This should apply most strongly to the later education of a machine arising from a child-machine of well-tried design (or program). This is in clear contrast with normal procedure when using a machine to do computations: one’s object is then to have a clear mental picture of the state of the machine at each moment in the computation. This object can only be achieved with a struggle. The view that “the machine can only do what we know how to order it to do”,³ appears strange in face of this. Most of the programs which we can put into the machine will result in its doing something that we cannot make sense of at all, or which we regard as completely random behavior. Intelligent behavior presumably consists in a departure from the completely disciplined behavior involved in computation, but a rather slight one, which does not give rise to random behavior, or to pointless repetitive loops. Another important result of preparing our machine for its part in the imitation game by a process of teaching and learning is that “human fallibility” is likely to be admitted in a rather natural way, that is, without special “coaching”. (The reader should reconcile this with the point of view on p. 113.) Processes that are learned do not produce a hundred percent certainty of result; if they did they could not be unlearned.

在学习机中加入随机元素可能是明智之举（见第 105 页）。当我们寻找某个问题的解决方案时，随机元素非常有用。例如，假设我们想找到一个介于 50 和 200 之间的数字，该数字等于其数字之和的平方，我们可以从 51 开始，然后尝试 52，一直到我们得到一个有效的数字。或者，我们可以随机选择数字，直到我们得到一个好的数字。这种方法的优点是不需要跟踪已经尝试过的值，但缺点是可能会尝试同一个值两次；但如果有多个解决方案，这并不重要。系统方法的缺点是，在必须首先调查的区域中可能会有一个巨大的没有任何解决方案的块。现在，学习过程可以看作是寻找一种能够满足老师（或其他标准）的行为形式。由于可能存在大量令人满意的解决方案，因此随机方法似乎比系统方法更好。应该注意的是，它被用于类似的进化过程。但在那里，系统方法是不可能的。如何才能跟踪已经尝试过的不同基因组合，以避免再次尝试它们？

It is probably wise to include a random element in a learning machine (see p. 105). A random element is rather useful when we are searching for a solution of some problem. Suppose for instance we wanted to find a number between 50 and 200 which was equal to the square of the sum of its digits, we might start at 51 then try 52 and go on until we got a number that worked. Alternatively we might choose numbers at random until we got a good one. This method has the advantage that it is unnecessary to keep track of the values that have been tried, but the disadvantage that one may try the same one twice; but this is not very important if there are several solutions. The systematic method has the disadvantage that there may be an enormous block without any solutions in the region which has to be investigated first. Now the learning process may be regarded as a search for a form of behavior which will satisfy the teacher (or some other criterion). Since there is probably a very large number of satisfactory solutions, the random method seems to be better than the systematic. It should be noticed that it is used in the analogous process of evolution. But there the systematic method is not possible. How could one keep track of the different genetical combinations that had been tried, so as to avoid trying them again?

我们可能希望机器最终在所有纯智力领域与人类竞争。但从哪个领域开始最好呢？即使是这个也很难决定。许多人认为，像下棋这样的非常抽象的活动是最好的。也可以坚持认为，最好为机器提供金钱可以买到的最好的感官，然后教它理解和说英语。这个过程可以遵循对孩子的正常教学。事物会被指出和命名，等等。同样，我不知道正确的答案是什么，但我认为两种方法都应该尝试一下。

We may hope that machines will eventually compete with men in all purely intellectual fields. But which are the best ones to start with? Even this is a difficult decision. Many people think that a very abstract activity, like the playing of chess, would be best. It can also be maintained that it is best to provide the machine with the best sense organs that money can buy, and then teach it to understand and speak English. This process could follow the normal teaching of a child. Things would be pointed out and named, and so on. Again I do not know what the right answer is, but I think both approaches should be tried.

我们只能看到前方很短的距离，但我们可以看到有很多事情需要做。

We can only see a short distance ahead, but we can see plenty there that needs to be done.

第 6 章注释

Notes for Chapter 6

1.   1.这种观点可能有点异端邪说。圣托马斯·阿奎那（《神学大全》，罗素 1945 年著，第 458 页）说，上帝不能使人没有灵魂。但这可能并不是对他权力的真正限制，而只是因为人的灵魂是不朽的，因此是不可毁灭的。

2.   1.  Possibly this view is heretical. St. Thomas Aquinas (Summa Theologica, quoted in Russell, 1945, p. 458) states that God cannot make a man to have no soul. But this may not be a real restriction on His powers, but only a result of the fact that men’s souls are immortal, and therefore indestructible.

3.   2.或者说，我们的儿童机器“被编程”将在数字计算机中被编程。但逻辑系统不需要学习。

4.   2.  Or rather “programmed in” for our child-machine will be programmed in a digital computer. But the logical system will not have to be learned.

5.   3.比较洛夫莱斯夫人的陈述（第 114 页），其中没有包含“仅”一词。

6.   3.  Compare Lady Lovelace’s statement (p. 114), which does not contain the word “only”.

7.1 智能行为

7.1  Intelligent behaviour

这篇论文是关于人工智能的科学。不幸的是，人工智能的技术吸引了所有人的注意力。公众可能会认为人工智能就是那些令人惊叹的应用，智能的和自主的。我们这个领域的人士都知道，对于许多应用来说，“智能”一词只不过是一个流行词（就像“红色美味苹果”中的“美味”一词一样）。除了考虑许多可能有益的人工智能应用外，我们经常对人工智能技术的潜在滥用（在武器等领域）抱有严重的疑虑。

This paper is about the science of AI. Unfortunately, it is the technology of AI that gets all the attention. The general public could be forgiven for thinking that AI is just about all those whiz-bang applications, smart this and autonomous that. Those of us in the field know that for many applications, the term “intelligent” is no more than a buzzword (like the term “delicious” in “red delicious apples”). And along with the many possibly beneficial AI applications under consideration, we often have serious misgivings about the potential misuse of AI technology (in areas like weaponry).

但人工智能不仅仅是一门技术。我们中的许多人并不是受到目前正在考虑的任何人工智能应用的激励，而是受到科学事业的激励，试图了解我们周围的世界。不同的科学有不同的主题，而人工智能是从计算角度研究智能行为的。还有什么比这更令人着迷的呢？人脑是一个了不起的东西，也许是我们所知的宇宙中最复杂的物体。但更了不起的是人脑所能做的事情。我们最好的智能行为远远超出了我们有权期待纯物理行为所能达到的范围。事实上，人工智能科学的首要问题是：物理的东西（比如人）如何可能真正做到X？

But AI is more than just technology. Many of us are motivated not by any of the AI applications currently being considered, but by the scientific enterprise, the attempt to understand the world around us. Different sciences have different subject matters, and AI is the study of intelligent behaviour in computational terms. What could be more fascinating? The human brain is a remarkable thing, perhaps the single most complex object we know of in the universe. But even more remarkable is what a human brain is capable of doing. Our intelligent behaviour at its best goes well beyond what we have any right to expect to emerge out of purely physical matter. Indeed, the overarching question for the science of AI is: How is it possible for something physical (like people, for instance) to actually do X?

其中X是智能行为的众多实例之一。这需要与一个相关问题进行对比：我们能否设计一个计算机系统来做一些模糊的X式的事情？

where X is one of the many instances of intelligent behaviour. This needs to be contrasted with a related question: Can we engineer a computer system to do something that is vaguely X-ish?

关于这一点我们后面还会有更多讨论。

about which we will have much more to say later.

请注意，人工智能科学研究的是智能行为，而不是谁或什么产生了这种行为。例如，它研究的是自然语言理解，而不是自然语言理解者。这就是人工智能与人类研究（神经科学、心理学、认知科学、进化生物学等）截然不同的地方。

Note that the science of AI studies intelligent behaviour, not who or what is producing the behaviour. It studies natural language understanding, for instance, not natural language understanders. This is what makes AI quite different from the study of people (in neuroscience, psychology, cognitive science, evolutionary biology, and so on).

我们关心的是哪种行为？不同的研究人员自然会关注不同的方面。行为可能依赖于感知或运动技能，也可能不依赖于感知或运动技能。行为可能包括学习，也可能不包括学习。行为可能基于情绪反应或社交互动，也可能不基于情绪反应或社交互动。一些研究人员主要关注的是各种动物的智能行为，比如在房间中找到所需物品的能力。而另一些人则只关注人类的行为，比如下棋的能力。（这两组人有时会发生方法论上的争论，前者认为，除非我们了解人类行为的基本形式，否则我们无法理解人类行为，而后者则认为科学根本不是这样运作的。在游戏的这个阶段，真的没有理由采取教条主义的立场。）

What sort of behaviour do we care about? Different researchers will quite naturally focus on different aspects. The behaviour may or may not depend on perceptual or motor skills. It may or may not include learning. It may or may not be grounded in emotional responses, or in social interactions. For some researchers, the main concern is intelligent behaviour seen in a variety of animals, like the ability to find a desired object in a room. For others, the focus is on behaviour seen in humans only, like the ability to play chess. (These two groups sometimes engage in methodological disputes, with the former arguing that we cannot expect to understand human behaviour until we understand its more basic forms, and the latter responding that this is not how science works at all. At this stage of the game, there is really no reason to take a doctrinaire position one way or another.)

7.2 行为测试

7.2  Behavioural tests

给定某种形式的智能行为，我们如何知道人工智能研究人员讲述的计算故事实际上解释了该行为。答案可以追溯到图灵：如果计算解释能够生成无法与人类行为长期区分的行为，那么它就是足够的。

Given some form of intelligent behaviour, how do we know that the computational story told by AI researchers actually explains the behaviour. The answer, going all the way back to Turing, is this: a computational account is adequate if it is able to generate behaviour that cannot be distinguished over the long haul from the behaviour produced by people.

这当然让人想起著名的图灵测试（图灵，1950 年；本卷第 6 章）。我们设想一个询问者与两个参与者（一个人和一台计算机）通过电传打字机进行长时间的对话。对话自然、流畅，可以讨论任何话题。如果无论对话有多长，询问者都无法分辨出两个参与者中的哪一个是人，则计算机通过了图灵测试。

This, of course, harks back to the famous Turing Test (Turing, 1950; Chapter 6 of this volume). We imagine an extended conversation over a teletype between an interrogator and two participants, a person and a computer. The conversation is natural, free-flowing, and about any topic whatsoever. The computer is said to pass the Turing Test if no matter how long the conversation, the interrogator cannot tell which of the two participants is the person.

在我看来，图灵在这一切中的观点是：“智能”、“思考”、“理解”等术语太过模糊和情绪化，不值得争论。如果我们坚持在科学背景下使用它们，我们应该愿意说，一个能够通过适当行为测试的程序与人一样具有所讨论的属性。改编电影人物阿甘正传的名言“愚蠢就是愚蠢的行为”，我们可以想象图灵说“聪明就是聪明的行为”。在我看来，这是一个非常明智的立场，我曾在其他地方为它辩护过（Levesque，2009）。

Turing’s point in all this, it seems to me, is this: Terms like “intelligent,” “thinking,” “understanding,” and the like are much too vague and emotionally charged to be worth arguing about. If we insist on using them in a scientific context at all, we should be willing to say that a program that can pass a suitable behavioural test has the property in question as much as the person. Adapting the dictum of the movie character Forest Gump who said “Stupid is as stupid does,” we can imagine Turing saying “Intelligent is as intelligent does.” This is a very sensible position, it seems to me, and I have defended it elsewhere (Levesque, 2009).

7.3 Winograd 模式问题

7.3  Winograd schema questions

我们的方法可以通过一个示例问题来最好地说明：⁵

Our approach is best illustrated with an example question:⁵

Winograd 模式问题是一个二元选择题，具有以下属性：

A Winograd schema question is a binary-choice question with these properties:

• 问题中提到了两方（均为男性、女性、物体或群体）。
• Two parties are mentioned in the question (both are males, females, objects, or groups).
• 代词用于指代其中一方（“他”，“她”，“它”或“他们”，根据当事人的说法）。
• A pronoun is used to refer to one of them (“he,” “she,” “it,” or “they,” according to the parties).
• 问题始终是相同的：代词的所指是什么？
• The question is always the same: what is the referent of the pronoun?
• 在幕后，该模式有两个特殊词。模式中有一个可用任一词填充的插槽。正确答案取决于选择哪个特殊词。
• Behind the scenes, there are two special words for the schema. There is a slot in the schema that can be filled by either word. The correct answer depends on which special word is chosen.

在上面，使用的特殊词是“given”，另一个词是“received”。因此，每个 Winograd 模式实际上都会生成两个非常相似的问题：

In the above, the special word used is “given,” and the other word is “received.” So each Winograd schema actually generates two very similar questions:

和

and

这两个问题之间只有一个词的差异，这有助于防止人们使用最低劣的伎俩来回答它们。

It is this one-word difference between the two questions that helps guard against using the cheapest of tricks on them.

这里还有一些其他的例子。第一个甚至适合小孩子：

Here are some additional examples. The first is one that is suitable even for young children:

在这种情况下，使用的特殊词是“小”，另一个词是“大”。这是 Terry Winograd (1972a) 的原始示例，该模式以他的名字命名：

In this case, the special word used is “small” and the other word is “big.” Here is the original example due to Terry Winograd (1972a) for whom the schema is named:

这里的特殊词是“害怕”，替代词是“提倡”。

Here the special word is “feared” and the alternative word is “advocated.”

只要稍加注意，就可以想出一些能锻炼不同专业知识的 Winograd 模式问题。以下是有关某些材料的示例：

With a bit of care, it is possible to come up with Winograd schema questions that exercise different kinds of expertise. Here is an example concerning certain materials:

特殊词是“泡沫聚苯乙烯”，替代词是“钢”。这个测试解决问题的能力：

The special word is “styrofoam” and the alternative is “steel.” This one tests for problem-solving skill:

特殊词是“below”，替代词是“above”。此示例测试可视化能力：

The special word is “below” and the alternative is “above.” This example tests for an ability to visualize:

使用的特殊词是“高尔夫球手”，另一个词是“狗”。

The special word used is “golfers” and the other is “dogs.”

当然，这种形式的问题并不适用于这里。可以构造太“简单”的问题，例如：

Of course not just any question in this form will do the job here. It is possible to construct questions that are too “easy,” like this one:

问题是，这个问题可以用以下技巧来回答：忽略给定的句子，并检查哪两个词更频繁地同时出现（根据谷歌的说法）：“racecar”与“fast”或“school bus”与“fast”。问题也可能太“难”，比如这个：

The problem is that this question can be answered using the following trick: ignore the given sentence, and check which two words co-occur more frequently (according to Google, say): “racecar” with “fast” or “school bus” with “fast.” Questions can also be too “hard,” like this one:

问题是，当使用“快乐”变体时，这个问题就变得模棱两可了。弗兰克可能因为他是赢家而感到快乐，也可能因为比尔是赢家而感到快乐。关于这些问题和其他问题的进一步讨论可以在 Levesque 等人 (2012) 的文章中找到。

The problem is that this question is ambiguous when the “happy” variant is used. Frank could plausibly be happy because he is the winner or because Bill is. Further discussion on these and other issues can be found in Levesque et al. (2012).

7.3.1 新的测试

7.3.1  A new test

现在可以制定图灵测试的替代方案。可以将一组预先测试过的 Winograd 模式隐藏在库中。6 ^Winograd模式测试涉及询问以下几个问题，对错误答案给予严厉惩罚（以阻止猜测）。测试可以以完全自动化的方式进行管理和评分：

It is now possible to formulate an alternative to the Turing Test. A collection of pre-tested Winograd schemas can be hidden in a library.⁶ A Winograd Schema Test involves asking a number of these questions with a strong penalty for wrong answers (to preclude guessing). A test can be administered and graded in a fully automated way:

1. 选择N（例如N =25）个合适的问题（就词汇量、专业知识等而言）；

1. select N (e.g., N = 25) questions that are suitable (with respect to vocabulary, expertise, etc.);

2. 随机使用问题中的一个特殊词语；

2. randomly use one of the special words in the question;

3. 向受试者展示测试，并获取N 个二进制答复；

3. present the test to the subject, and obtain the N binary replies;

考试的最终成绩是

The final grade for the test is

其中k表示猜测的惩罚（例如，k = 5）。这里的主要主张是，英语能力正常的成年人将轻松通过测试。因此，如果我们想要产生与人类无法区分的行为，我们需要想出一个也可以通过测试的程序。

where k codes the penalty for guessing (e.g., k = 5). The main claim here is that normally-abled English-speaking adults will pass the test easily. So, if we want to produce behaviour that is indistinguishable from that of people, we will need to come up with a program that can also pass the test.

总结一下：关于图灵测试，我们同意图灵的观点，即实质性问题是计算机程序是否可以实现某种智能行为。但图灵所提倡的自由形式对话可能不是正式测试的最佳载体，因为它允许狡猾的受试者躲在嬉戏、语言技巧和预设回答的烟幕后面。我们的立场是，基于 Winograd 模式问题的替代测试不太容易被滥用，但显然比进行合作对话（例如，关于图灵想象的十四行诗）对智力的要求要低得多。

To summarize: With respect to the Turing Test, we agree with Turing that the substantive question is whether or not a certain intelligent behaviour can be achieved by a computer program. But a free-form conversation as advocated by Turing may not be the best vehicle for a formal test, as it allows a cagey subject to hide behind a smokescreen of playfulness, verbal tricks, and canned responses. Our position is that an alternative test based on Winograd schema questions is less subject to abuse, though clearly much less demanding intellectually than engaging in a cooperative conversation (about sonnets, for example, as imagined by Turing).

7.4 通过测试

7.4  Passing the test

计算机程序需要什么才能通过 Winograd 模式测试？我的感觉是，我们可以在以下方面取得很大进展：

What would it take for a computer program to pass a Winograd schema test. My feeling is that we can go quite some distance with the following:

对于上述问题，可得出以下结论：

For the question above, this gives the following:

• – x不适合y + x相对于 来说太小了 。
• – x does not fit in y + x is so small vs.
• – x不适合y + y太小
• – x does not fit in y + y is so small

这种“大数据”方法是一个很好的技巧，但不幸的是，它仍然太廉价了。除其他外，它忽略了R和P之间的联系。考虑一下：

This “big data” approach is an excellent trick, but unfortunately, it is still too cheap. Among other things, it ignores the connective between R and P. Consider this:

请注意，这里的R和P与之前相同，尽管这次的答案一定不同。

Note that the R and P here would be the same as before, even though the answer must be different this time.

现在考虑以下示例：

Now consider the following example:

这里任何R和P之间的关系显然要复杂得多。

Here the relationship between any R and P is clearly much more complex.

那么，我们能从中得出什么结论呢？我们只是需要更多的技巧吗？

So what do we conclude from this? Do we simply need a bigger bag of tricks?

7.5 两个科学障碍

7.5  Two scientific hurdles

我并不想暗示，有了麦卡锡的激进思想，一切就都一帆风顺了。一个值得思考的问题是，为什么 55 年后，我们在智能行为科学方面取得的成果如此之少。我认为，答案是它留下了一些未解决的重大问题。

I do not want to suggest that with McCarthy’s radical idea on board, it is all smooth sailing from here. A good question to ask is why, after 55 years, we have so little to show for it regarding the science of intelligent behaviour. The answer, I believe, is that it leaves some major issues unresolved.

我在 IJCAI-85 上的《计算机与思想》讲座（Levesque，1986）部分是对当时非常流行的“知识就是力量”口号的回应。即使在当时，这一切对我来说似乎都太肤浅了。我的感觉是，如果知识不能以合适的符号形式获得，或者不能以易于处理的方式应用，它就不是力量。这指出了麦卡锡主义面临的两个重大障碍：

My Computers and Thought Lecture at IJCAI-85 (Levesque, 1986) was in part a reaction to the “Knowledge is Power” slogan which was quite in vogue at the time. It all seemed too facile to me, even back then. My sense was that knowledge was not power if it could not be acquired in a suitable symbolic form, or if it could not be applied in a tractable way. These point to two significant hurdles faced by the McCarthy approach:

我认为，这两个障碍对人工智能科学来说就像宇宙加速膨胀对天体物理学一样严重和具有挑战性。55 年后，我们很可能会想知道人工智能研究人员是否能够克服它们。

I believe that these two hurdles are as serious and as challenging to the science of AI as an accelerating universe is to astrophysics. After 55 years, we might well wonder if an AI researcher will ever be able to overcome them.

生命短暂（“上市时间”更短），许多人工智能研究人员回归到不那么激进的方法（例如，更多基于生物学，更像统计力学）来关注那些看似知识密集程度较低的行为（例如，识别手写数字、在人群中跟踪面孔、在崎岖地形上行走），这也许并不奇怪。而且结果非常棒！

Life being short (and “time to market” even shorter), it is perhaps not surprising that many AI researchers have returned to less radical methods (e.g., more biologically-based, more like statistical mechanics) to focus on behaviours that are seemingly less knowledge-intensive (e.g., recognizing handwritten digits, following faces in a crowd, walking over rough terrain). And the results have been terrific!

但这些出色的结果不应该让我们否认。我们最好的行为确实包括知识密集型活动，例如参与自然对话或回答 Winograd 模式问题。我希望我们中有足够多的人专注于这种智能行为，以便继续取得进展。

But these terrific results should not put us into denial. Our best behaviour does include knowledge-intensive activities such as participating in natural conversations, or responding to Winograd schema questions. It is my hope that enough of us stay focused on this sort of intelligent behaviour to allow progress to continue here as well.

这需要付出艰苦的努力！我认为，期望解决方案会自发地从一些一般原则中产生，而不需要我们付出任何真正的努力，这是不合理的。例如，我认为我们永远无法构建一个小型计算机程序，给它一个摄像头和一个麦克风，或者把它放到网上，然后期望它自己获得它所需要的一切。

This will require hard work! I think it is unreasonable to expect solutions to emerge spontaneously out of a few general principles, obviating any real effort on our parts. For example, I do not think we will ever be able to build a small computer program, give it a camera and a microphone or put it on the web, and expect it to acquire what it needs all by itself.

所以工作会很辛苦。但在我看来，这更像是攀登高山，而不是铲除车道上的积雪。工作确实很辛苦，但也是一次令人兴奋的冒险！

So the work will be hard. But to my way of thinking, it will be more like scaling a mountain than shoveling a driveway. Hard work, yes, but an exhilarating adventure!

第 7 章注释

Notes for Chapter 7

1.   1.对于那些不知道的人来说，障碍赛是一种赛马比赛，与通常的赛马比赛类似，但马匹必须跳过赛道上的许多树篱。所以它就像马的跨栏赛。

2.   1.  For those who do not know, a steeplechase is a horse race, similar to the usual ones, but where the horses must jump over a number of hedges on the racetrack. So it is like hurdles for horses.

3.   2.请参阅 Brian Christian (2011) 的书，其中有关于在洛布纳竞赛中扮演人类的有趣描述。

4.   2.  See the book by Brian Christian (2011) for an interesting account of what it was like to play the human in a Loebner contest.

5.   3. ELIZA 程序（Weizenbaum，1966）是解决这个问题的一个好起点。

6.   3.  The ELIZA program (Weizenbaum, 1966) is a good place to start on that issue.

7.   4. www.trueknowledge.com 上的程序似乎以这种方式运行

8.   4.  The program at www.trueknowledge.com appears to work this way

9.   5.本节主要摘自 Levesque 等人 (2012)。感谢 Ernie Davis 和 Leora Morgenstern 的贡献。

10.   5.  This section is drawn mainly from Levesque et al. (2012). I thank Ernie Davis and Leora Morgenstern for their contribution.

11.   6.例如，请参阅http://www.cs.nyu.edu/faculty/davise/papers/WS.html上的合集。​​​

12.   6.  See, for example, the collection at http://www.cs.nyu.edu/faculty/davise/papers/WS.html.

13.   7.确实，在 Winograd 模式问题上尝试使用Watson会很有趣：类别是“代词指称”，线索是“琼向苏珊表示感谢，感谢她给予的所有帮助”，期望以问题形式给出的答案是“苏珊是谁？”

14.   7.  Indeed, it would be good fun to try Watson on Winograd schema questions: the category is “Pronoun referents,” the clue is “Joan made sure to thank Susan for all the help she had given,” and the desired answer in the form of a question is “Who is Susan?”

15.   8.这些事实摘自维基百科有关泡沫塑料的页面。

16.   8.  These facts were lifted from the Wikipedia page for styrofoam.

8.1 人工智能

8.1  Artificial intelligence

人工智能领域的最终目标往往定义不明确，且存在争议。一些研究人员的目标是模拟人类认知，另一些研究人员的目标是创造智能而不考虑人类特征，还有一些研究人员的目标是创造有用的人工制品而不考虑智能的抽象概念。

AI is a field whose ultimate goal has often been somewhat ill-defined and subject to dispute. Some researchers aim to emulate human cognition, others aim at the creation of intelligence without concern for human characteristics, and still others aim to create useful artifacts without concern for abstract notions of intelligence.

这种多样性并不一定是坏事，因为每种方法都会揭示新的想法，并为其他方法提供养料。但有人可能会说，由于哲学家们痛恨定义真空，许多关于人工智能可行性的破坏性和不实之词的争论都是关于我们作为人工智能研究人员不认同的人工智能定义。

This variety is not necessarily a bad thing, since each approach uncovers new ideas and provides fertilization to the others. But one can argue that, since philosophers abhor a definitional vacuum, many of the damaging and ill-informed debates about the feasibility of AI have been about definitions of AI to which we as AI researchers do not subscribe.

我研究人工智能的动机是将智能视为系统的一般属性，而不是人类的特定属性，并将其理解为智能。我相信这对于整个领域来说是一个合适的目标，当然也包括创造有用的人工制品——既是副产品，也是技术发展的焦点和驱动力。然而，这种“创造智能”观点的困难在于，它假定我们对智能有一些富有成效的概念。认知科学家可以说“看，我的模型正确地预测了人类认知的实验观察”，人工制品开发人员可以说“看，我的系统正在拯救生命/创造巨额资金”，但我们中很少有人会对论文说“看，我的系统是智能的”感到满意。这一困难进一步加剧了理论框架的需要，以便我们能够自信地设计复杂的系统并借鉴他人的成果。“智能”必须有一个可以直接与系统的输入、结构和输出相关的定义。这样的定义也必须是通用的。否则，人工智能就会陷入各种各样的领域——下棋的智能、车辆控制的智能、医疗诊断的智能。

My own motivation for studying AI is to create and understand intelligence as a general property of systems, rather than as a specific attribute of humans. I believe this to be an appropriate goal for the field as a whole, and it certainly includes the creation of useful artifacts—both as a spin-off and as a focus and driving force for technological development. The difficulty with this “creation of intelligence” view, however, is that it presupposes that we have some productive notion of what intelligence is. Cognitive scientists can say “Look, my model correctly predicted this experimental observation of human cognition”, and artifact developers can say “Look, my system is saving lives/megabucks”, but few of us are happy with papers saying “Look, my system is intelligent”. This difficulty is compounded further by the need for theoretical scaffolding to allow us to design complex systems with confidence and to build on the results of others. “Intelligent” must be given a definition that can be related directly to the system’s input, structure, and output. Such a definition must also be general. Otherwise, AI subsides into a smorgasbord of fields—intelligence as chess playing, intelligence as vehicle control, intelligence as medical diagnosis.

在本文中，我将概述此类定义在人工智能和相关学科的发展历程中的发展。我将把每个定义视为一个谓词P，据说可以用来描述智能系统。对于每个P ，我将讨论“看，我的系统是P ”这一陈述是否有趣，至少有时是正确的，以及P系统的研究将导致什么样的研究和技术发展。

In this paper, I shall outline the development of such definitions over the history of AI and related disciplines. I shall examine each definition as a predicate P that can be applied, supposedly, to characterize systems that are intelligent. For each P, I shall discuss whether the statement “Look, my system is P” is interesting and at least sometimes true, and the sort of research and technological development to which the study of P-systems leads.

首先，我认为智能与成功行为的能力密切相关，即所谓的“基于代理”的人工智能观点。智能的正式定义如下：

I shall begin with the idea that intelligence is strongly related to the capacity for successful behaviour—the so-called “agent-based” view of AI. The candidates for formal definitions of intelligence are as follows:

• P ₁：完全理性，或者根据现有信息产生最大程度成功行为的能力。
• P₁: Perfect rationality, or the capacity to generate maximally successful behaviour given the available information.
• P2：计算理性，或者根据最初可用的信息计算出完全理性决策的原则能力_。
• P₂: Calculative rationality, or the in-principle capacity to compute the perfectly rational decision given the initially available information.
• P ₃：元层次理性，或者在动作必须由计算选择的约束下，选择计算序列加动作的最佳组合的能力。
• P₃: Metalevel rationality, or the capacity to select the optimal combination of computation-sequence-plus-action, under the constraint that the action must be selected by the computation.
• P ₄：有界优化，或在可用信息和计算资源的情况下产生最大成功行为的能力。
• P₄: Bounded optimality, or the capacity to generate maximally successful behaviour given the available information and computational resources.

我将详细阐述这四个定义，并描述迄今为止在这方面取得的一些成果。然后，我将描述计算理性和有界最优性方面的正在进行和未来的工作。

All four definitions will be fleshed out in detail, and I will describe some results that have been obtained so far along these lines. Then I will describe ongoing and future work under the headings of calculative rationality and bounded optimality.

我认为，在这些候选方案中，有界最优性最接近满足人工智能研究的需求。这种说法始终存在危险，即接受它会导致“过早数学化”，这种情况的特点是，结果越来越技术化，而与原始问题（在人工智能的情况下，即创造智能的问题）的关系却越来越小。有界最优性研究是否适合替代智能研究？我希望表明P _{4 （有界最优性）比}P ₁到P ₃更合适，因为它是一个具有真实和理想解决方案的真实问题，也因为它满足了关于智能本质的一些基本直觉。一些关于智能的重要问题只能在有界最优性或其相关框架内制定和回答。然而，只有时间才能告诉我们，有界最优性研究（也许经过进一步改进）是否能够产生足够的理论框架来支持人工智能的重大实际进展。

I shall be arguing that, of these candidates, bounded optimality comes closest to meeting the needs of AI research. There is always a danger, in this sort of claim, that its acceptance can lead to “premature mathematization”, a condition characterized by increasingly technical results that have increasingly little to do with the original problem—in the case of AI, the problem of creating intelligence. Is research on bounded optimality a suitable stand-in for research on intelligence? I hope to show that P₄, bounded optimality, is more suitable than P₁ through P₃ because it is a real problem with real and desirable solutions, and also because it satisfies some essential intuitions about the nature of intelligence. Some important questions about intelligence can only be formulated and answered within the framework of bounded optimality or some relative thereof. Only time will tell, however, whether bounded optimality research, perhaps with additional refinements, can generate enough theoretical scaffolding to support significant practical progress in AI.

8.2 代理

8.2  Agents

直到最近，人们普遍将人工智能定义为“心智能力”或“智能系统”的计算研究，对各种类型进行分类，然后就此打住。这并没有提供太多指导。相反，人们可以将人工智能定义为设计做正确事情的系统的问题。现在我们只需要对“正确”进行定义。

Until fairly recently, it was common to define AI as the computational study of “mental faculties” or “intelligent systems”, catalogue various kinds, and leave it at that. This does not provide much guidance. Instead, one can define AI as the problem of designing systems that do the right thing. Now we just need a definition for “right”.

这种方法涉及将智能实体视为代理，即感知其环境并对其采取行动的系统。正式地说，代理是通过从感知序列到代理实例化的操作的映射来定义的。让O成为集合代理可以随时观察到的感知，A是代理可以在外部世界执行的可能动作集（包括什么也不做）。因此，代理函数 f : O ^* → A定义了代理在任何情况下的行为方式。首先重要的是代理做了什么，而不一定是它想了什么，甚至它是否思考。这种最初拒绝考虑对代理内部运作的进一步限制（例如，它应该进行逻辑推理）有三点好处：首先，它允许我们将诸如规划和推理之类的“认知能力”视为为找到正确的事情而发生的；其次，它包含而不是排除系统可以在没有这种认知能力的情况下做正确的事情的立场（Agre 和 Chapman，1987；Brooks，1989）；第三，它允许更自由地考虑子系统的各种规范、边界和互连。

This approach involves considering the intelligent entity as an agent, that is to say a system that senses its environment and acts upon it. Formally speaking, an agent is defined by the mapping from percept sequences to actions that the agent instantiates. Let O be the set of percepts that the agent can observe at any instant, and A be the set of possible actions the agent can carry out in the external world (including the action of doing nothing). Thus the agent function f: O^* → A defines how an agent behaves under all circumstances. What counts in the first instance is what the agent does, not necessarily what it thinks, or even whether it thinks at all. This initial refusal to consider further constraints on the internal workings of the agent (such as that it should reason logically, for example) helps in three ways: first, it allows us to view such “cognitive faculties” as planning and reasoning as occurring in the service of finding the right thing to do; second, it encompasses rather than excludes the position that systems can do the right thing without such cognitive faculties (Agre and Chapman, 1987; Brooks, 1989); third, it allows more freedom to consider various specifications, boundaries, and interconnections of subsystems.

基于代理的人工智能观点已迅速从“情境性”和“嵌入性”研讨会转移到主流教科书（Dean 等人，1995 年；Russell 和 Norvig，1995 年）和《新闻周刊》的流行语中。宽泛地说，理性代理是指其行为从代理所拥有的信息及其目标（或设计其执行的任务）的角度来看有意义的代理。理性是行为的一种属性，它并不指定（尽管它确实限制了）选择行为的过程。这是 Simon（1956 年）强调的一点，他创造了实质性理性和程序理性这两个术语来描述做出什么决策的问题和如何做出决策的问题之间的区别。 Rod Brooks 1991 年的计算机与思维讲座题为“没有理性的智能”（另见 Brooks，1991），强调了这样一个事实：推理（也许）是代理的派生属性，可能是也可能不是实现理性行为的良好实施方案。证明许多人工智能研究人员认为理所当然的认知结构的合理性并非易事。

The agent-based view of AI has moved quickly from workshops on “situatedness” and “embeddedness” to mainstream textbooks (Dean et al., 1995; Russell and Norvig, 1995) and buzzwords in Newsweek. Rational agents, loosely speaking, are agents whose actions make sense from the point of view of the information possessed by the agent and its goals (or the task for which it was designed). Rationality is a property of actions and does not specify—although it does constrain—the process by which the actions are selected. This was a point emphasized by Simon (1956), who coined the terms substantive rationality and procedural rationality to describe the difference between the question of what decision to make and the question of how to make it. That Rod Brooks’1991 Computers and Thought lecture was titled “Intelligence without Reason” (see also Brooks, 1991) emphasizes the fact that reasoning is (perhaps) a derived property of agents that might, or might not, be a good implementation scheme to achieve rational behaviour. Justifying the cognitive structures that many AI researchers take for granted is not an easy problem.

基于代理的智能观点的另一个结果是，它使人工智能面临来自其他领域的竞争，这些领域传统上将嵌入式代理视为一个自然的研究主题。控制理论是其中最重要的，但进化编程和进化生物学本身也有想法可以贡献^。1

One other consequence of the agent-based view of intelligence is that it opens AI up to competition from other fields that have traditionally looked on the embedded agent as a natural topic of study. Control theory is foremost among these, but evolutionary programming and indeed evolutionary biology itself also have ideas to contribute.¹

代理视图的流行也帮助该领域朝着解决实际问题的方向发展，避免了布鲁克斯所说的“幻觉”问题，即当子系统的脆弱性被智能人类提供输入并解释其输出所掩盖时出现的“幻觉”问题。

The prevalence of the agent view has also helped the field move towards solving real problems, avoiding what Brooks calls the “hallucination” problem that arises when the fragility of a subsystem is masked by having an intelligent human providing input to it and interpreting its outputs.

8.3 完全理性

8.3  Perfect rationality

完全理性会限制代理的行为，以在现有信息下提供最大的成功期望。我们可以将这一概念扩展如下（见图8.1）。定义的基本输入是代理将在其中运行的环境类E和评估以下序列的绩效指标U代理通过这些状态来驱动实际环境。令V ( f, E , U ) 表示代理函数f在环境类E中获得的根据U 的预期值，其中（现在）我们将假设E元素的概率分布。然后，完全理性的代理由代理函数f _opt定义，即

Perfect rationality constrains an agent’s actions to provide the maximum expectation of success given the information available. We can expand this notion as follows (see figure 8.1). The fundamental inputs to the definition are the environment class E in which the agent is to operate and the performance measure U which evaluates the sequence of states through which the agent drives the actual environment. Let V(f, E, U) denote the expected value according to U obtained by an agent function f in environment class E, where (for now) we will assume a probability distribution over elements of E. Then a perfectly rational agent is defined by an agent function f_opt such that

这只是一种奇特的说法，即最佳代理会尽其所能。关键在于，完全理性的行为是E和U的一个明确定义的函数，我将其称为任务环境。下面将讨论计算此函数的问题。

This is just a fancy way of saying that the best agent does the best it can. The point is that perfectly rational behaviour is a well-defined function of E and U, which I will call the task environment. The problem of computing this function is addressed below.

完美理性在人工智能中的理论作用在 Newell 的论文《知识层面》（Newell，1982 年）中得到了很好的描述。人工智能系统的知识层面分析依赖于完美理性的假设。它可以用来确定任何可能系统的性能上限，即确定一个完全理性的代理在拥有相同知识的情况下会做什么。

The theoretical role of perfect rationality within AI is well-described by Newell’s paper on the Knowledge Level (Newell, 1982). Knowledge-level analysis of AI systems relies on an assumption of perfect rationality. It can be used to establish an upper bound on the performance of any possible system, by establishing what a perfectly rational agent would do given the same knowledge.

尽管完全理性的主体所拥有的知识决定了它在实现目标后将采取的行动，但知识从何而来的问题尚未得到很好的理解。也就是说，我们需要理解理性学习和理性行动。从理性的逻辑观点来看，学习几乎没有受到关注——事实上，纽厄尔的分析排除了知识层面的学习。从决策理论的观点来看，贝叶斯更新为理性学习提供了一个模型，但这将问题推回到了先验（Carnap，1950）。理性先验的问题，特别是对于表达性表示语言的问题，仍未得到解决。

Although the knowledge that a perfectly rational agent has determines the actions that it will take given its goals, the question of where the knowledge comes from is not well understood. That is, we need to understand rational learning as well as rational action. In the logical view of rationality, learning has received almost no attention—indeed, Newell’s analysis precludes learning at the knowledge level. In the decision-theoretic view, Bayesian updating provides a model for rational learning, but this pushes the question back to the prior (Carnap, 1950). The question of rational priors, particularly for expressive representation languages, remains unsettled.

完美理性的另一个方面是缺乏一套适当的技术来指定效用函数。在经济学中，许多结果都是通过把整体效用分解成可以以各种方式组合的属性而得出的（Keeney 等人，1976 年），但这种方法在人工智能领域几乎没有取得任何进展（但参见 Bacchus 和 Grove，1995 年；Wellman，1985 年）。我们也不太清楚如何指定随时间变化的效用，尽管这个问题经常被提出，但我们对目标和效用之间的关系并没有令人满意的理解。

Another aspect of perfect rationality that is lacking is the development of a suitable body of techniques for the specification of utility functions. In economics, many results have been derived on the decomposition of overall utility into attributes that can be combined in various ways (Keeney et al., 1976), yet such methods have made few inroads into AI (but see Bacchus and Grove, 1995; Wellman, 1985). We also have little idea how to specify utility over time, and although the question has been raised often, we do not have a satisfactory understanding of the relationship between goals and utility.

完全理性代理的优点在于，如果你手头有一个代理，你会更喜欢它而不是其他代理。此外，如果你是一名经济学家，你可以证明由它们组成的经济体的良好结果；如果你想设计分布式智能系统，假设每个代理都具有完全理性，交互机制的设计就会容易得多。缺点是，完全理性理论不提供代理内部设计的分析：一个完全理性的代理与另一个完全理性的代理一样好。正如西蒙所指出的，真正糟糕的是，完全理性的代理并不存在。物理机制需要时间来处理信息和选择动作，因此真实代理的行为无法立即反映环境的变化，并且通常不是最优的。

The good thing about perfectly rational agents is that if you have one handy, you prefer it to any other agent. Furthermore, if you are an economist you can prove nice results about economies populated by them; and if you want to design distributed intelligent systems, assuming perfect rationality on the part of each agent makes the design of the interaction mechanisms much easier. The bad thing is that the theory of perfect rationality does not provide for the analysis of the internal design of the agent: one perfectly rational agent is as good as another. The really bad thing, as pointed out by Simon, is that perfectly rational agents do not exist. Physical mechanisms take time to process information and select actions, hence the behaviour of real agents cannot immediately reflect changes in the environment and will generally be suboptimal.

8.4 计算合理性

8.4  Calculative rationality

在讨论计算理性之前，有必要介绍一下代理函数和代理程序之间的区别。在人工智能中，代理被实现为一个程序，我将其称为l，它在一台机器上运行，我将其称为M。代理程序接收当前感知作为输入，但也有内部状态，以某种形式反映先前的感知。它在选择动作后输出动作。从外部来看，代理的行为由选定的动作和不动作（或机器生成的任何默认动作）组成。

Before discussing calculative rationality, it is necessary to introduce a distinction between the agent function and the agent program, In AI, an agent is implemented as a program, which I shall call l, running on a machine, which I shall call M. An agent program receives as input the current percept, but also has internal state that reflects, in some form, the previous percepts. It outputs actions when they have been selected. From the outside, the behaviour of the agent consists of the selected actions interspersed with inaction (or whatever default actions the machine generates).

计算理性表现在程序中，如果执行速度无限快，将产生完全理性的行为。与完全理性不同，计算理性是许多实际程序可以满足的要求。与完全理性不同的是，计算理性不一定是一种理想的特性。例如，一个计算理性的国际象棋程序会选择“正确”的举动，但这样做可能需要 10 ⁵⁰倍的时间。

Calculative rationality is displayed by programs that, if executed infinitely fast, would result in perfectly rational behaviour. Unlike perfect rationality, calculative rationality is a requirement that can be fulfilled by many real programs. Also unlike perfect rationality, calculative rationality is not necessarily a desirable property. For example, a calculatively rational chess program will choose the “right” move, but may take 10⁵⁰ times too long to do so.

尽管如此，追求计算理性一直是人工智能理论研究的主要活动。在该领域的早期阶段，重要的是要专注于“认识论充分性”而不是“启发式充分性”——即原则上的能力而不是实践上的能力。2^计算理性一直是逻辑和决策理论传统的支柱。在逻辑传统中，绩效衡量标准接受在所有情况下实现指定目标的行为，并拒绝任何其他行为。因此，纽厄尔（1982）将理性行为定义为那些保证实现代理的目标之一。逻辑规划系统（例如使用情境演算的定理证明器）满足此定义下的计算理性条件。在决策理论传统中，计算理性代理的设计主要在人工智能之外进行 - 例如，在随机最优控制理论（Kumar and Varaiya，1986）中。表示通常非常贫乏（基于状态而不是句子），可解问题要么非常小，要么非常专业。在人工智能中，概率网络或信念网络的发展为代理设计开辟了许多新的可能性，在许多情况下使表示复杂性呈指数级降低。基于影响图的系统（添加了动作和值节点的概率网络）满足决策理论版本的计算理性。

The pursuit of calculative rationality has nonetheless been the main activity of theoretically well-founded research in AI. In the early stages of the field, it was important to concentrate on “epistemological adequacy” before “heuristic adequacy”—that is, capability in principle rather than in practice.² Calculative rationality has been the mainstay of both the logical and the decision-theoretic traditions. In the logical tradition, the performance measure accepts behaviours that achieve the specified goal in all cases and rejects any others. Thus Newell (1982) defines rational actions as those that are guaranteed to achieve one of the agent’s goals. Logical planning systems, such as theorem-provers using situation calculus, satisfy the conditions of calculative rationality under this definition. In the decision-theoretic tradition, the design of calculatively rational agents has largely gone on outside AI—for example, in stochastic optimal control theory (Kumar and Varaiya, 1986). Representations have usually been very impoverished (state-based rather than sentential) and solvable problems have been either very small or very specialized. Within AI, the development of probabilistic networks or belief networks has opened up many new possibilities for agent design, providing in many cases an exponential reduction in representational complexity. Systems based on influence diagrams (probabilistic networks with action and value nodes added) satisfy the decision-theoretic version of calculative rationality.

实际上，无论是逻辑还是决策理论传统都无法避免计算理性要求所带来的决策问题的棘手性。一种应对措施是排除所处理的表示和推理任务中指数复杂性的来源，以便计算理性和完美理性相一致——至少，如果我们忽略多项式时间计算的小问题。这一立场在 1985 年 Hector Levesque（Levesque，1986 年；Levesque 和 Brachman，1987 年）和 1989 年 Henry Kautz 的两次精彩的计算机和思想讲座中得到了阐述。关于可处理子语言的随附研究结果可能最好被视为复杂性可能成为问题的迹象，而不是复杂性问题的解决方案。Doyle 和 Patil（1991 年）强烈反对限制表达能力的想法，他们指出，这也限制了在这种约束下设计的表示和推理服务的适用性^。3

In practice, neither the logical nor the decision-theoretic traditions can avoid the intractability of the decision problems posed by the requirement of calculative rationality. One response is to rule out sources of exponential complexity in the representations and reasoning tasks addressed, so that calculative and perfect rationality coincide—at least, if we ignore the little matter of polynomial-time computation. This position was expounded in two fascinating Computers and Thought lectures given by Hector Levesque in 1985 (Levesque, 1986; Levesque and Brachman, 1987) and by Henry Kautz in 1989. The accompanying research results on tractable sublanguages are perhaps best seen as indications of where complexity may be an issue rather than as a solution to the problem of complexity. The idea of restricting expressiveness was strongly opposed by Doyle and Patil (1991), who pointed out that it also restricts the applicability of the representation and inference services designed under such constraints.³

在分布式人工智能领域，系统设计者可以控制每个代理环境中涉及与其他代理协商的部分。因此，控制复杂性的一种可能方法是限制协商问题，以便轻松做出最佳决策。例如，可以使用克拉克税机制来确保每个代理的最佳策略只是如实陈述其偏好（Ephrati 和 Rosenschein，1991 年）。当然，这种方法不一定会导致代理集合的最佳行为；也不能解决与环境其他部分交互的复杂性问题。

In the area of distributed AI, the system designer has control over that part of each agent’s environment that involves negotiations with other agents. Thus, one possible way to control complexity is to constrain the negotiation problem so that optimal decisions can be made easily. For example, the Clarke Tax mechanism can be used to ensure that the best policy for each agent is simply to state its preferences truthfully (Ephrati and Rosenschein, 1991). Of course, this approach does not necessarily result in optimal behaviour by the ensemble of agents; nor does it solve the problem of complexity in interacting with the rest of the environment.

应对复杂性最常见的方法是使用各种加速技术和近似值，以期获得合理的行为。人工智能已经开发出一套非常强大的方法来降低复杂性，包括将状态表示分解为句子形式；环境模型的稀疏表示（如 STRIPS 运算符）；解决方案分解方法，如偏序规划和抽象；用于强化学习的价值函数的近似、参数化表示；编译（chucking、宏运算符、EBL 等）；以及元级控制的应用。虽然其中一些方法可以保留最优性的保证，并且对于结构良好的中等规模问题有效，但智能代理不可避免地无法在所有情况下都采取理性行动。自人工智能诞生之初，这种观察就很常见。然而，选择次优行动的系统本身并不属于计算理性，我们需要更好的理论来理解它们。

The most common response to complexity has been to use various speedup techniques and approximations in the hope of getting reasonable behaviour. AI has developed a very powerful armoury of methods for reducing complexity, including the decomposition of state representations into sentential form; sparse representations of environment models (as in STRIPS operators); solution decomposition methods such as partial-order planning and abstraction; approximate, parameterized representations of value functions for reinforcement learning; compilation (chucking, macro-operators, EBL, etc.); and the application of metalevel control. Although some of these methods can retain guarantees of optimality and are effective for moderately large problems that are well structured, it is inevitable that intelligent agents will be unable to act rationally in all circumstances. This observation has been a commonplace since the very beginning of AI. Yet systems that select suboptimal actions fall outside calculative rationality per se, and we need a better theory to understand them.

8.5 元层次理性

8.5  Metalevel rationality

元级理性，也被 IJ Good (1971) 称为第二类理性，基于在计算成本和决策质量之间寻找最佳平衡的理念。尽管 Good 从未精确阐述过第二类理性的概念——他将其定义为“考虑审议成本的情况下预期效用的最大化”——但很明显，其目的是利用某种元级架构来实现这种平衡。元级架构是一种智能代理的设计理念，它将代理划分为两个（或更多）概念部分。对象级执行与应用领域有关的计算——例如，预测物理动作的结果、计算某些状态的效用等等。元级是第二个决策过程，其应用领域包括对象级计算本身以及它们影响的计算对象和状态。元推理在人工智能领域有着悠久的历史，至少可以追溯到 20 世纪 70 年代初（有关历史详细信息，请参阅 Russell 和 Wefald，1991a）。人们还可以将选择性搜索方法和修剪策略视为体现有关追求特定对象级搜索操作的可取性的元级专业知识。

Metalevel rationality, also called Type II rationality by I.J Good (1971), is based on the idea of finding an optimal tradeoff between computational costs and decision quality. Although Good never made his concept of Type II rationality very precise—he defines it as “the maximization of expected utility taking into account deliberation costs—it is clear that the aim was to take advantage of some sort of metalevel architecture to implement this tradeoff. Metalevel architecture is a design philosophy for intelligent agents that divides the agent into two (or more) notional parts. The object level carries out computations concerned with the application domain—for example, projecting the results of physical actions, computing the utility of certain states, and so on. The metalevel is a second decision-making process whose application domain consists of the object-level computations themselves and the computational objects and states that they affect. Metareasoning has a long history in AI, going back at least to the early 1970s (see Russell and Wefald, 1991a for historical details). One can also view selective search methods and pruning strategies as embodying metalevel expertise concerning the desirability of pursuing particular object-level search operations.

理性元推理理论将元层次能够“正确思考”的直觉形式化。其基本思想是，对象层次的计算是有成本（时间流逝）和收益（决策质量的提高）的行为。理性元层次根据预期效用选择计算。理性元推理的前身是信息价值理论（Howard，1966 年）——该概念认为，通过模拟信息请求的每个可能结果下将遵循的决策过程，可以计算出获取额外信息的决策理论价值，从而估计这些结果平均的决策质量预期改进。与信息收集类似，计算过程的应用似乎起源于 Matheson（1968 年）。在人工智能领域，Horvitz (1987a,b)、Breese 和 Fehling (1990) 以及 Russell 和 Wefald (1989, 1991a,b) 都展示了计算价值的思想如何解决实时决策的基本问题。

The theory of rational metareasoning formalizes the intuition that the metalevel can “do the right thinking.” The basic idea is that object-level computations are actions with costs (the passage of time) and benefits (improvements in decision quality). A rational metalevel selects computations according to their expected utility. Rational metareasoning has as a precursor the theory of information value (Howard, 1966)—the notion that one can calculate the decision-theoretic value of acquiring an additional piece of information by simulating the decision process that would be followed given each possible outcome of the information request, thereby estimating the expected improvement in decision quality averaged over those outcomes. The application to computational processes, by analogy to information-gathering, seems to have originated with Matheson (1968). In AI, Horvitz (1987a,b), Breese and Fehling (1990), and Russell and Wefald (1989, 1991a,b) all showed how the idea of value of computation could solve the basic problems of real-time decision-making.

Eric Wefald 所做的工作特别关注搜索算法，其中对象级计算将各种行动结果的预测扩展到未来。例如，在国际象棋程序中，每个对象级计算都会扩展游戏树的一个叶节点。然后，元级问题是选择要扩展的节点并在适当的点终止搜索。此类系统中元推理的主要问题是，计算的局部效应不会直接转化为改进的决策，因为还有一个复杂的过程，即将叶节点的局部效应传播回根节点和移动选择。事实证明，可以根据“局部努力”和“传播函数”找到计算值的通用公式，这样就可以为任何特定的对象级系统（例如极小最大传播）实例化该公式，并在运行时高效编译和执行。该方法已用于双人游戏、具有机会节点的双人游戏和单代理搜索。在每种情况下，相同的通用元推理方案都会比传统的、高度工程化的算法产生大约一个数量级的效率提升。

The work done with Eric Wefald looked in particular at search algorithms, in which the object-level computations extend projections of the results of various courses of actions further info the future. For example, in chess programs, each object-level computation expands a leaf node of the game tree. The metalevel problem is then to select nodes for expansion and to terminate search at the appropriate point. The principal problem with metareasoning in such systems is that the local effects of the computations do not directly translate into improved decisions, because there is also a complex process of propagating the local effects at the leaf back to the root and the move choice. It turns out that a general formula for the value of computation can be found in terms of the “local efforts” and the “propagation function”, such that the formula can be instantiated for any particular object-level system (such as minimax propagation), compiled, and executed efficiently at runtime. This method was implemented for two-player games, two-player games with chance nodes, and single-agent search. In each case, the same general metareasoning scheme resulted in efficiency improvements of roughly an order of magnitude over traditional, highly-engineered algorithms.

另一类常见的元推理问题出现在随时（Dean and Boddy，1988）或灵活（Horvitz，1987a）算法中，这些算法旨在返回结果，其质量随分配给计算的时间量而变化。最简单的元推理类型是将单个算法的决策质量预期提升（以性能配置文件衡量）与时间成本（Simon，1955）进行权衡。如果性能配置文件的二阶导数为负，则贪婪终止条件是最佳的。如果希望从随时组件构建复杂的实时系统，则会出现更复杂的问题。首先，必须确保组成系统的可中断性- 即确保整个系统能够稳健地响应即时的输出需求。解决方案是交错执行所有组件，为每个组件分配时间，以便系统每个完整迭代改进周期的总时间在每次迭代时加倍。通过这种方式，我们可以构建一个复杂系统，该系统可以处理任意和意外的实时需求，就好像它提前知道可用的确切时间一样，速度只会受到很小的（≤4）常数因子损失（Russell and Zilberstein，1991）。其次，必须在组件之间优化分配可用的计算，以最大化总输出质量。虽然这在一般情况下是 NP 难的，但当组件的调用图是树形结构时，它可以在与程序大小成线性关系的时间内求解（Zilberstein and Russell，1996）。虽然这些结果是在具有明确定义性能配置文件的任何时间算法的相对干净的背景下得出的，但我们有理由期待，复杂系统中鲁棒实时决策的一般问题可以在实践中得到处理。

Another general class of metareasoning problems arises with anytime (Dean and Boddy, 1988) or flexible (Horvitz, 1987a) algorithms, which are algorithms designed to return results whose quality varies with the amount of time allocated to computation. The simplest type of metareasoning trades off the expected increase in decision quality for a single algorithm, as measured by a performance profile, against the cost of time (Simon, 1955). A greedy termination condition is optimal if the second derivative of the performance profile is negative. More complex problems arise if one wishes to build complex real-time systems from anytime components. First, one has to ensure the interruptibility of the composed system—that is, to ensure that the system as a whole can respond robustly to immediate demands for output. The solution is to interleave the execution of all the components, allocating time to each component so that the total time for each complete iterative improvement cycle of the system doubles at each iteration. In this way, we can construct a complex system that can handle arbitrary and unexpected real-time demands exactly as if it knew the exact time available in advance, with just a small ( ≤ 4) constant factor penalty in speed (Russell and Zilberstein, 1991). Second, one has to allocate the available computation optimally among the components to maximize the total output quality. Although this is NP-hard for the general case, it can be solved in time linear in program size when the call graph of the components is tree-structured (Zilberstein and Russell, 1996). Although these results are derived in the relatively clean context of anytime algorithms with well-defined performance profiles, there is reason to expect that the general problem of robust real-time decision-making in complex systems can be handled in practice.

过去几年，关于元知识和元推理的性质出现了一场有趣的争论。TEIRESIAS（Davis，1980）提出了这样一种观点，即明确的、特定领域的元知识是专家系统创建的一个重要方面。因此，元知识是一种“额外”领域知识，超出了对象级领域知识，人们必须将其添加到 AI 系统中才能使其正常工作。另一方面，在上述理性元推理的研究中，很明显，描述计算效果的元理论与领域无关（Ginsberg 和 Geddis，1991；Russell 和 Wefald，1991a）。原则上，不需要额外的领域知识来评估计算的好处。在实践中，从第一原理进行元推理的成本可能非常高。为了避免这种情况，可以将特定领域的元级分析结果汇编成领域特定的元知识，或者直接从经验中学习此类知识（见 Russell 和 Wefald，1991a，第 6 章，以及 Minton，1996 年）。这种对新兴“计算专业知识”的看法带来了对智能的根本洞察——即，从一种有趣的意义上讲，算法并不是人工智能系统的必要组成部分。相反，我们可以想象一个理性引导的计算的一般过程，它与环境属性相互作用，产生越来越高效的决策。在我看来，这种思维方式巧妙地解决了人工智能的一个主要难题：如果人工智能需要的是极其狡猾和超高效的算法，远远超出了当前计算机科学家的最佳努力，那么进化（以及机器学习又将如何）实现这一目标？

Over the last few years, an interesting debate has emerged concerning the nature of metaknowledge and metareasoning. TEIRESIAS (Davis, 1980) established the idea that explicit, domain-specific metaknowledge was an important aspect of expert system creation. Thus, metaknowledge is a sort of “extra” domain knowledge, over and above the object-level domain knowledge, that one has to add to an AI system to get it to work well. On the other hand, in the work on rational metareasoning described above, it is clear that the metatheory describing the effects of computations is domain-independent (Ginsberg and Geddis, 1991; Russell and Wefald, 1991a). In principle, no additional domain knowledge is needed to assess the benefits of a computation. In practice, metareasoning from first principles can be very expensive. To avoid this, the results of metalevel analysis for particular domains can be compiled into domain-specific metaknowledge, or such knowledge can be learned directly from experience (see Russell and Wefald, 1991a, Chapter 6, and Minton, 1996). This view of emerging “computational expertise” leads to a fundamental insight into intelligence—namely, that there is an interesting sense in which algorithms are not a necessary part of AI systems. Instead, one can imagine a general process of rationally guided computation interacting with properties of the environment to produce more and more efficient decision making. To my mind, this way of thinking finesses one major puzzle of AI: if what is required for AI is incredibly devious and superbly efficient algorithms far surpassing the current best efforts of computer scientists, how did evolution (and how will machine learning) ever get there?

在理性元推理领域，仍然存在一些重大的未解决的问题。一个明显的困难是，迄今为止，几乎所有系统都采用了一种短视的策略——在元级进行贪婪的深度搜索。显然，计算序列的最佳选择问题至少与底层对象级问题一样棘手。尽管如此，必须考虑序列，因为在某些情况下，计算的价值可能在进一步计算之前不会明显改善决策质量。这表明强化学习技术可能是有效的，特别是因为计算的“奖励函数”（即决策质量的改善）很容易在事后用于元级。其他可能的研究领域包括为更复杂的系统创建有效的元级控制器，例如抽象层次规划器、混合架构等。

Significant open problems remain in the area of rational metareasoning. One obvious difficulty is that almost all systems to date have adopted a myopic strategy—a greedy, depth-one search at the metalevel. Obviously, the problem of optimal selection of computation sequences is at least as intractable as the underlying object-level problem. Nonetheless, sequences must be considered because in some cases the value of a computation may not be apparent as an improvement in decision quality until further computations have been done. This suggests that techniques from reinforcement learning could be effective, especially as the “reward function” for computation—that is, the improvement in decision quality—is easily available to the metalevel post hoc. Other possible areas for research include the creation of effective metalevel controllers for more complex systems such as abstraction hierarchy planners, hybrid architectures, and so on.

尽管理性的元推理似乎是应对复杂性的有用工具，但元层次理性作为资源受限代理的正式框架的概念似乎站不住脚。原因是，由于元推理成本高昂，因此无法以最佳方式进行。对象层次理性的历史在元层次上重演：元层次上的完美理性是无法实现的，元层次上的计算理性是无用的。因此，必须在元层次计算中进行时间/最优性权衡，例如上述的短视近似。然而，在元层次理性的框架内，没有办法确定时间和元层次决策质量之间的适当权衡。任何通过元层次这样做的尝试只会导致概念倒退。此外，在某些环境中，最有效的代理设计完全有可能根本不进行元推理，而只是对情况做出反应。这些考虑表明，正确的方法是走出代理；避免对代理做出的个别决策进行微观管理。这是有界优化所采用的方法。

Although rational metareasoning seems to be a useful tool in coping with complexity, the concept of metalevel rationality as a formal framework for resource-bounded agents does not seem to hold water. The reason is that, since metareasoning is expensive, it cannot be carried out optimally. The history of object-level rationality has repeated itself at the metalevel: perfect rationality at the metalevel is unattainable and calculative rationality at the metalevel is useless. Therefore, a time/optimality tradeoff has to be made for metalevel computations, as for example with the myopic approximation mentioned above. Within the framework of metalevel rationality, however, there is no way to identify the appropriate tradeoff of time for metalevel decision quality. Any attempt to do so via a metametalevel simply results in a conceptual regress. Furthermore, it is entirely possible that in some environments, the most effective agent design will do no metareasoning at all, but will simply respond to circumstances. These considerations suggest that the right approach is to step outside the agent, as it were; to refrain from micromanaging the individual decisions made by the agent. This is the approach taken in bounded optimality.

8.6 有界最优性

8.6  Bounded optimality

完全理性和元层次理性的困难源于对事物（行为、计算）施加的约束，而这些约束不是代理设计者直接控制的。如果没有真正的代理能够满足规范，那么指定行为或计算是理性的也是毫无意义的。设计者控制着程序。在 Russell 和 Subramanian (1995) 的著作中，引入了给定机器的可行性概念来描述可由该机器上运行的某些代理程序实现的所有代理函数集。这有点类似于可计算性的概念，但要严格得多，因为它将程序在具有有限速度的形式机器模型上的操作与代理生成的实际时间行为联系起来。

The difficulties with perfect rationality and metalevel rationality arise from the imposition of constraints on things (actions, computations) that the agent designer does not directly control. Specifying that actions or computations be rational is of no use if no real agents can fulfill the specification. The designer controls the program. In Russell and Subramanian (1995), the notion of feasibility for a given machine is introduced to describe the set of all agent functions that can be implemented by some agent program running on that machine. This is somewhat analogous to the idea of computability, but is much stricter because it relates the operation of a program on a formal machine model with finite speed to the actual temporal behaviour generated by the agent.

鉴于这一观点，人们立即想到最优可行行为是一个有趣的概念，并想到找到生成它的程序。假设我们将Agent ( l, M ) 定义为由在机器M上运行的程序l实现的代理函数。那么有界最优程序l _opt定义为

Given this view, one is led immediately to the idea that optimal feasible behaviour is an interesting notion, and to the idea of finding the program that generates it. Suppose we define Agent(l, M) to be the agent function implemented by the program l running on machine M. Then the bounded optimal program l_opt is defined by

其中 ℒ _M是可以在M上运行的所有程序的有限集。这就是P ₄。有界最优性。

where ℒ_M is the finite set of all programs that can be run on M. This is P₄. bounded optimality.

在人工智能领域，有限优化的理念在 20 世纪 80 年代末期的几个对资源有限理性这一一般主题感兴趣的讨论小组中流传，尤其是罗克韦尔（由迈克尔·费林组织）和斯坦福（由迈克尔·布拉特曼组织）的小组。“有限优化”这一术语似乎由埃里克·霍维茨 (1987b) 首创，他将其非正式地定义为“在给定一组关于预期问题和资源约束的假设的情况下，计算效用的优化”。

In AI, the idea of bounded optimality floated around among several discussion groups interested in the general topic of resource-bounded rationality in the late 1980s, particularly those at Rockwell (organized by Michael Fehling) and Stanford (organized by Michael Bratman). The term “bounded optimality” seems to have been originated by Eric Horvitz (1987b), who defined it informally as “the optimization of computational utility given a set of assumptions about expected problems and constraints on resources”.

博弈论中也出现了类似的想法，即从考虑游戏中的最佳决策转向考虑最佳决策程序。这会导致不同的结果，因为它限制了每个代理对另一个代理进行无限模拟的能力，而后者也在对第一个代理进行无限模拟，依此类推。即使是可计算性的要求也会产生很大的不同（Megiddo 和 Wigderson，1986 年）。玩家复杂性的界限也已成为人们密切关注的话题。Papadimitriou 和 Yannakakis（1994 年）已经证明，如果每个代理都是一个有限自动机，并且其状态数小于轮数的指数，则迭代囚徒困境博弈中存在协作均衡。这本质上是一个有界的优化结果，其中界限是空间而不是计算速度。

Similar ideas have also surfaced recently in game theory, where there has been a shift from consideration of optimal decisions in games to a consideration of optimal decision-making programs. This leads to different results because it limits the ability of each agent to do unlimited simulation of the other, who is also doing unlimited simulation of the first, and so on. Even the requirement of computability makes a significant difference (Megiddo and Wigderson, 1986). Bounds on the complexity of players have also become a topic of intense interest. Papadimitriou and Yannakakis (1994) have shown that a collaborative equilibrium exists for the iterated Prisoner’s Dilemma game if each agent is a finite automaton with a number of states that is less than exponential in the number of rounds. This is essentially a bounded optimality result, where the bound is on space rather than speed of computation.

哲学对理性的定义也经历了逐渐的演变。从行为功利主义（个人行为的理性）转向规则功利主义（普遍的理性）。行动的政策。Cherniak (1986) 和 Harman (1983) 广泛讨论了政策对有限代理可行的要求。在“道德急救手册”(Dennett, 1986) 中可以找到与有界最优概念通常一致的哲学提议。Dennett 明确讨论了在可行决策程序空间内达到最优的想法，并以哲学系的博士录取程序为例。他指出，有界最优录取程序可能有些混乱，可能没有明显的“最优”标志——事实上，招生委员会可能会继续对其进行修改，因为有界最优系统可能无法识别其自身的有界最优性。

Philosophy has also seen a gradual evolution in the definition of rationality. There has been a shift from consideration of act utilitarianism—the rationality of individual acts—to rule utilitarianism, or the rationality of general policies for acting. The requirement that policies be feasible for limited agents was discussed extensively by Cherniak (1986) and Harman (1983). A philosophical proposal generally consistent with the notion of bounded optimality can be found in the “Moral First Aid Manual” (Dennett, 1986). Dennett explicitly discusses the idea of reaching an optimum within the space of feasible decision procedures, using as an example the Ph.D. admissions procedure of a philosophy department. He points out that the bounded optimal admissions procedure may be somewhat messy and may have no obvious hallmark of “optimality”—in fact, the admissions committee may continue to tinker with it since bounded optimal systems may have no way to recognize their own bounded optimality.

在与 Devika Subramanian 合作的过程中，有界最优性的一般思想被置于正式环境中，以便人们可以开始推导出有界最优程序的严格结果。这涉及在代理、程序、机器、环境和时间之间建立完全指定的关系。我们发现这本身就是一个非常有价值的练习。例如，“民间人工智能”的“实时环境”和“截止日期”概念最终的定义与我们最初想象的完全不同。在此基础上，研究了一种非常简单的机器架构，其中程序由固定执行时间和决策质量的决策程序组成。在“随机截止日期”环境中，事实证明，通过按顺序运行多个程序直到中断所获得的效用通常高于任何单个决策程序所能获得的效用。也就是说，在开始进行更复杂的计算之前，最好先准备一个“快速而粗略”的答案，以防后者无法及时完成。

In work with Devika Subramanian, the general idea of bounded optimality has been placed in a formal setting so that one can begin to derive rigorous results on bounded optimal programs. This involves setting up completely specified relationships among agents, programs, machines, environments, and time. We found this to be a very valuable exercise in itself. For example, the “folk AI” notions of “real-time environments” and “deadlines” ended up with definitions rather different than those we had initially imagined. From this foundation, a very simple machine architecture was investigated in which the program consists of decision procedures of fixed execution time and decision quality. In a “stochastic deadline” environment, it turns out that the utility attained by running several procedures in sequence until interrupted is often higher than that attainable by any single decision procedure. That is, it is often better first to prepare a “quick and dirty” answer before embarking on more involved calculations in case the latter do not finish in time.

这些结果的有趣之处在于，除了它们作为有界最优性的非平凡证明的价值之外，它们还以一种简单的方式展示了我认为有界最优代理的主要特征：有限机器内追求最优的压力会导致更复杂的程序结构。直观地说，在复杂环境中进行有效的决策需要一种提供各种可能的计算选项的软件架构，以便在大多数情况下，代理至少有一些可用的计算，从而显著提高决策质量。

The interesting aspect of these results, beyond their value as a demonstration of nontrivial proofs of bounded optimality, is that they exhibit in a simple way what I believe to be a major feature of bounded optimal agents: the fact that the pressure towards optimality within a finite machine results in more complex program structures. Intuitively, efficient decision-making in a complex environment requires a software architecture that offers a wide variety of possible computational options, so that in most situations the agent has at least some computations available that provide a significant increase in decision quality.

对上述有界最优性基本模型的一个可能的反对意见是，解决方案对于环境或机器的微小变化并不稳健。这反过来会导致分析复杂系统设计时遇到困难。理论计算机科学在描述算法的运行时间时面临同样的问题，因为计算步骤和描述指令集在最佳算法上会给出同样脆弱的结果。O() 符号就是为处理这个问题而开发的，它提供了一种更稳健的方式来描述复杂度，而不受机器速度和实现细节的影响。这种稳健性对于允许复杂性结果累积发展也是必不可少的。在 Russell 和 Subramanian (1995) 中，相应的概念是渐近有界最优性 (ABO)。与经典复杂性一样，我们可以定义平均情况和最坏情况 ABO，其中“情况”在这里表示环境。例如，最坏情况 ABO 定义如下：

One possible objection to the basic model of bounded optimality outlined above is that solutions are not robust with respect to small variations in the environment or the machine. This in turn would lead to difficulties in analysing complex system designs. Theoretical computer science faced the same problem in describing the running time of algorithms, because counting steps and describing instruction sets exactly gives the same kind of fragile results on optimal algorithms. The O() notation was developed to deal with this and provides a much more robust way to describe complexity that is independent of machine speeds and implementation details. This robustness is also essential in allowing complexity results to develop cumulatively. In Russell and Subramanian (1995), the corresponding notion is asymptotic bounded optimality (ABO). As with classical complexity, we can define both average-case and worst-case ABO, where “case” here means the environment. For example, worst-case ABO is defined as follows:

最坏情况渐近有界最优性。 代理程序 l 在时间上（或空间上）是 E 中M 上的最坏情况 ABO，当且仅当

Worst-case asymptotic bounded optimality. An agent program l is timewise (or space-wise) worst-case ABO in E on M iff

其中 kM 表示速度提高了 k 倍（或内存增加了 k 倍）的 M 版本，而 V ^* ( f, E , U, n )是对于E中所有复杂度为 n 的E 而言，V ( f, E , U )的最小值。

where kM denotes a version of M speeded up by a factor k (or with k times more memory) and V^*(f, E, U, n) is the minimum value of V(f, E, U) for all E in E of complexity n.

用英语来说，这意味着如果该程序只需要一台更快（更大）的机器来实现与所有环境中的任何其他程序一样好的最坏情况行为，那么该程序基本上是沿着正确的路线运行的。

In English, this means that the program is basically along the right lines if it just needs a faster (larger) machine to have worst-case behaviour as good as that of any other program in all environments.

对有界优化理念的另一个可能的反对意见是，它只是将元级理性的棘手计算负担从代理的元级转移到设计者的对象级。当然，有人可能会说，设计师现在必须离线解决所有在线时难以解决的元级优化问题。这种说法并非毫无道理——事实上，如果代理设计问题变得简单，那将是令人惊讶的。然而，这两个问题之间存在显著差异，因为代理设计者可能正在为整个环境类创建代理，而假定的元级代理正在特定环境中工作。通过考虑排序算法的例子可以看出，这可以使设计师的问题更容易解决。对一万亿个元素的列表进行排序可能确实非常困难，但设计一个渐近最优的排序算法相对容易。事实上，这两项任务的难度是无关的。对于 BO 和 ABO 设计来说，这种不相关性仍然成立，但是 ABO 定义使其更加清晰。

Another possible objection to the idea of bounded optimality is that it simply shifts the intractable computational burden of metalevel rationality from the agent’s metalevel to the designer’s object level. Surely, one might argue, the designer now has to solve offline all the metalevel optimization problems that were intractable when online. This argument is not without merit—indeed, it would be surprising if the agent design problem turns out to be easy. There is however, a significant difference between the two problems, in that the agent designer is presumably creating an agent for an entire class of environments, whereas the putative metalevel agent is working in a specific environment. That this can make the problem easier for the designer can be seen by considering the example of sorting algorithms. It may be very difficult indeed to sort a list of a trillion elements, but it is relatively easy to design an asymptotically optimal algorithm for sorting. In fact, the difficulties of the two tasks are unrelated. The unrelatedness would still hold for BO as well as ABO design, but the ABO definitions make it a good deal clearer.

可以很容易地证明，最坏情况 ABO 是渐近最优算法的泛化，只需构建一个“经典环境”，经典算法在其中运行，并且如果输出正确，则算法行为的效用是运行时的递减正函数，否则为零。更一般环境中的代理可能需要权衡输出质量和时间，随着时间的推移生成多个输出，等等。为了说明 ABO 是一种有用的抽象，可以证明在某些限制下，可以使用 Russell 和 Zilberstein (1991) 的加倍构造构建通用ABO 程序，这些程序对于效用函数的任何时间变化都是 ABO。下面讨论了有界优化研究的进一步方向。

It can be shown easily that worst-case ABO is a generalization of asymptotically optimal algorithms, simply by constructing a “classical environment” in which classical algorithms operate and in which the utility of the algorithm’s behaviour is a decreasing positive function of runtime if the output is correct and zero otherwise. Agents in more general environments may need to trade off output quality for time, generate multiple outputs over time, and so on. As an illustration of how ABO is a useful abstraction, one can show that under certain restrictions one can construct universal ABO programs that are ABO for any time variation in the utility function, using the doubling construction from Russell and Zilberstein (1991). Further directions for bounded optimality research are discussed below.

8.7 该怎么办？

8.7  What is to be done?

本节将介绍一些研究活动，我希望这些研究活动有助于将有界最优性转变为 AI 系统设计的创新工具。不过，首先，我将介绍需要完成的计算合理系统方面的工作，以丰富代理程序的空间。

This section describes some of the research activities that will, I hope, help to turn bounded optimality into a creative tool for AI system design. First, however, I shall describe work on calculatively rational systems that needs to be done in order to enrich the space of agent programs.

8.8 总结

8.8  Summary

我已概述了一些正式的人工智能研究方向，这些研究以有界最优性作为人工智能系统的理想属性。这种对人工智能的看法似乎是从优化动作或计算到优化程序这一不可避免的哲学“转变”的逻辑结果。我曾提出，这种方法应该允许理论和实践人工智能研究之间产生协同作用，而这种协同作用是其他正式框架所无法提供的。同样，我相信它是创造智能这一非正式目标令人满意的正式对应物。特别是，它完全符合我们对真实智能代理需要复杂结构、相对较小的思维在广阔世界中面临的资源限制的重要性以及进化作为设计优化过程的运作的直觉。人们还可以说，有界最优性研究可能会更好地满足那些希望模仿人类智能的人的需求，因为它考虑到了计算资源的限制，而这些限制可能是人类表现出的大多数令人遗憾的偏离完美理性的原因。

I have outlined some directions for formally grounded AI research based on bounded optimality as the desired property of AI systems. This perspective on AI seems to be a logical consequence of the inevitable philosophical “move” from optimization over actions or computations to optimization over programs. I have suggested that such an approach should allow synergy between theoretical and practical AI research of a kind not afforded by other formal frameworks. In the same vein, I believe it is a satisfactory formal counterpart of the informal goal of creating intelligence. In particular, it is entirely consistent with our intuitions about the need for complex structure in real intelligent agents, the importance of the resource limitations faced by relatively tiny minds in large worlds, and the operation of evolution as a design optimization process. One can also argue that bounded optimality research is likely to satisfy better the needs of those who wish to emulate human intelligence, because it takes into account the limitations on computational resources that are presumably responsible for most of the regrettable deviation from perfect rationality exhibited by humans.

当然，有界最优性和它的近亲渐近最优性只不过是人们希望系统满足的正式定义的属性。现在判断 ABO 是否会为人工智能做与渐近复杂性为理论计算机科学所做的相同的工作还为时过早。设计中的创造力仍然是人工智能研究人员的特权。然而，有可能将设计过程系统化，并自动化使系统适应其计算资源和环境需求的过程。有界最优性的概念提供了一种确保适应过程“正确”的方法。

Bounded optimality and its asymptotic cousin are, of course, nothing but formally defined properties that one may want systems to satisfy. It is too early to tell whether ABO will do the same kind of work for AI that asymptotic complexity has done for theoretical computer science. Creativity in design is still the prerogative of AI researchers. It may, however be possible to systematize the design process somewhat and to automate the process of adapting a system to its computational resources and the demands of the environment. The concept of bounded optimality provides a way to make sure the adaptation process is “correct”.

我希望通过这些研究，最终能够开发出概念和数学工具来回答一些关于智能的基本问题。例如，为什么复杂的智能系统（似乎）具有声明性知识结构，并对其进行明确推理？这是将人工智能与其他代理设计学科区分开来的一个基本假设，但答案仍然未知。事实上，罗德·布鲁克斯、休伯特·德雷福斯和其他人断然否认了这一假设。很明显，需要某种类似有界最优代理设计理论的东西来回答这个问题。

My hope is that with these kinds of investigations, it will eventually be possible to develop the conceptual and mathematical tools to answer some basic questions about intelligence. For example, why do complex intelligent systems (appear to) have declarative knowledge structures over which they reason explicitly? This has been a fundamental assumption that distinguishes AI from other disciplines for agent design, yet the answer is still unknown. Indeed, Rod Brooks, Hubert Dreyfus, and others flatly deny the assumption. What is clear is that it will need something like a theory of bounded optimal agent design to answer this question.

我在这里讨论的大多数代理设计特征，包括使用陈述性知识，都是在“首先构建计算理性代理，然后加速它们”的标准方法中构思出来的。然而，人们有理由怀疑这种方法是否能让人工智能界发现通用智能所需的所有设计特征。原因是，任何可以想象的计算机都无法接近完美理性，即使是在中等复杂的环境中。完美理性，如果你愿意的话，是智能代理的“牛顿”定义，而现实世界是粒子加速器。很可能基于计算理性设计的改进的代理甚至无法达到在给定底层计算资源的情况下可能达到的性能水平。出于这个原因，我认为必须不要忽视那些乍一看似乎不符合“经典”计算理性框架的代理设计想法。相反，人们必须尝试了解相应建筑类别中有限的最优配置的潜力，并看看是否可以设计适当的适应机制来帮助实现这些配置。

Most of the agent design features that I have discussed here, including the use of declarative knowledge, have been conceived within the standard methodology of “first build calculatively rational agents and then speed them up”. Yet one can legitimately doubt that this methodology will enable the AI community to discover all the design features needed for general intelligence. The reason is that no conceivable computer will ever be remotely close to approximating perfect rationality for even moderately complex environments. Perfect rationality is, if you like, a “Newtonian” definition for intelligent agents whereas the real world is a particle accelerator. It may well be the case that agents based on improvements to calculatively rational designs are not even close to achieving the level of performance that is potentially achievable given the underlying computational resources. For this reason, I believe it is imperative not to dismiss ideas for agent designs that do not seem at first glance to fit into the “classical” calculatively rational framework. Instead, one must attempt to understand the potential of the bounded optimal configurations within the corresponding architectural class, and to see if one can design the appropriate adaptation mechanisms that might help in realizing these configurations.

正如上一节所述，在制作更通用、更强大的“砖块”以构建更现实环境的人工智能系统方面还有很多工作要做，这些工作将为实现有界最优提供更多空间。从某种意义上说，在这个概念下，人工智能研究现在和它应该一直都是一样。

As mentioned in the previous section, there is also plenty of work to do in the area of making more general and more robust “bricks” from which to construct AI systems for more realistic environments, and such work will provide added scope for the achievement of bounded optimality. In a sense, under this conception AI research is the same now as it always should have been.

第 8 章注释

Notes for Chapter 8

1.   1.我认为这是一个非常积极的发展。人工智能是一个由其问题而不是其方法定义的领域。其主要见解——其中包括为决策服务而学习、使用和汇编显性知识——当然可以经受住来自其他领域的新方法的涌入。当其他领域同时接受从人工智能中获得的见解时，情况尤其如此。

2.   1.  I view this as a very positive development. AI is a field defined by its problems, not its methods. Its principal insights—among them the learning, use, and compilation of explicit knowledge in the service of decision making—can certainly withstand the influx of new methods from other fields. This is especially true when other fields are simultaneously embracing the insights derived within AI.

3.   2.这一决定是在计算机科学界正确理解计算难解性问题之前做出的，这或许并非巧合。

4.   2.  Perhaps not coincidentally, this decision was taken before the question of computational intractability was properly understood in computer science.

5.   3. Doyle 和 Patil (1991) 提出了“推理的理性管理”这一理念。表示系统“应设计为提供成本和质量各异的广泛服务”，并应考虑“系统用户所感知的[计算]成本和收益”。也就是说，他们提出了一种基于理性元推理的解决方案，如第 8.5 节所述。

6.   3.  Doyle and Patil (1991) propose instead the idea of “rational management of inference”. Representation systems “should be designed to offer a broad mix of services varying in cost and quality” and should take into account “the costs and benefits [of computations] as perceived by the system’s user”. That is, they suggest a solution based on rational metareasoning, as discussed in Section 8.5.

7.   4. Horvitz 和 Breese (1990) 也做出了类似的观察，他们研究的是对象级别受到严格限制的情况，因此元级别决策问题可以在恒定时间内得到解决。

8.   4.  A similar observation was made by Horvitz and Breese (1990) for cases where the object level is so restricted that the metalevel decision problem can be solved in constant time.

第 9 章注释

Notes for Chapter 9

1.   1.当然，这只是一种理想化的想法；关于相信什么（无论是非主观的还是其他的）的决定通常无法充分利用现有数据。然而，这种考虑并不影响本文的观点，即此类决定必须对来自许多不同来源的信息敏感。

2.   1.  This is, of course, an idealization; decisions about what to believe (subdoxastically or otherwise) do not, in general, succeed in making the optimal use of the available data. This consideration does not, however, affect the present point, which is just that such decisions must, of necessity, be sensitive to information from many different sources.

3.   2.这种论证方式背后有一个假设，读者可能不愿意承认：垂直能力之间的接口机制必须是计算性的，而不是人们所说的机械性的。关于语言如何与感知相连的旧观点（例如，感知是图像，而文字是其关联物）隐含地否定了这一假设。然而，在我看来，任何认真思考在决定（例如）如何表达我们所看到的东西时必须涉及什么的人都会接受这种观点的合理性，即所涉及的心理过程必须是计算性的，并且具有巨大的复杂性。

4.   2.  There is an assumption underlying this line of argument which the reader may not wish to grant: that the mechanisms that interface between vertical faculties have to be computational rather than, as one might say, merely mechanical. Old views of how language connects with perception (e.g., percepts are pictures and words are their associates) implicitly deny this assumption. It seems to me, however, that anyone who thinks seriously about what must be involved in deciding (e.g.) how to say what we see will accept the plausibility of the view that the mental processes that are implicated must be both computational and of formidable complexity.

5.   3 .  编者注：这些属性在（Fodor，1983）的第三部分中定义。Fodor 认为输入系统是特定于领域的、强制性的，只向中央系统提供有限的信息，信息封装，具有“浅层”输出，并具有典型的故障模式。

6.   3.  Editors’ note; these properties are defined in Part III of (Fodor, 1983). Fodor argues that input systems are domain-specific, mandatory, provide only limited information to the central system, are informationally encapsulated, have ‘shallow’ outputs, and have characteristic breakdown patterns.

7.   4.更强有力的版本认为，每个理论陈述在逻辑上必须等同于一些（有限的？）观察陈述的结合。有关此文献的详细评论，请参阅 Glymour (1980)。Glymour 对 Quineian 确认论的某些方面持异议，但原因与我们在此讨论的无关。

8.   4.  Stronger versions had it that each theory statement must be logically equivalent to some (finite?) conjunction of observation statements. For a sophisticated review of this literature, see Glymour (1980). Glymour takes exception to some aspects of the Quineian account of confirmation, but not for reasons that need concern us here.

9.   5.人们经常提出（例如参见 McCarthy，1980），能够处理框架问题的逻辑必须是“非单调的”。（粗略地说，当增加新的公设不会减少先前可推导的定理集时，逻辑是单调的；否则，逻辑是非单调的。）关键在于，新的信念不会直接添加到旧信念中；相反，旧信念会以各种方式改变以适应新的信念。然而，根据文中提出的框架问题的分析，这并不奇怪。因为，就此而言，框架问题与非演示确认问题没有区别，确认关系本身通常是非单调的。例如，新数据的可用性可能需要为无限多的先前接受的假设分配新的确认级别。因此，如果我们将确认系统视为形式化的，那么每当有新数据可用时，无限多的先前可推导出的形式为“H 的水平是 L”的公式可能会变成非定理。

10.   5.  It is often proposed (see, e.g., McCarthy, 1980) that a logic capable of coping with the frame problem will have to be ‘nonmonotonic’. (Roughly, a logic is monotonic when the addition of new postulates does not reduce the set of previously derivable theorems; nonmonotonic otherwise.) The point is that new beliefs don’t just get added on to the old set; rather, old beliefs are variously altered to accommodate the new ones. This is, however, hardly surprising on the analysis of the frame problem proposed in the text. For, on that account, the frame problem is not distinguishable from the problem of nondemonstrative confirmation, and confirmation relations are themselves typically nonmonotonic. For example, the availability of a new datum may necessitate the assignment of new confirmation levels to indefinitely many previously accepted hypotheses. Hence, if we think of the confirmation system as formalized, indefinitely many previously derivable formulas of the form ‘the level of H is L’ may become nontheorems whenever new data become available.

11.   6.由于框架问题和框架兼数据结构之间没有特别的关系，因此该领域的命名法非常令人困惑。

12.   6.  Since there is no particular relation between the frame problem and frames-cum-data structures, the nomenclature in this area could hardly be more confusing.

13.   7.定位之争当然并未随着 Gall 和 Flourens 的终结而结束。有关其相对较近的历史（自 Wernicke 以来）的简要概述，请参阅 Eggert (1977)。顺便说一句，Wernicke 固然在语言机制方面是坚定的定位主义者，但他坚持认为只有“基本功能……才能归属于特定区域。所有超出这些基本功能的过程（例如将各种感知综合为概念以及诸如思维和意识之类的复杂功能）都依赖于连接皮层不同区域的纤维束”（第 92 页），这一点颇有趣味。除了联想论之外，Wernicke 的图景与我们在此阐述的图景并无太大区别。

14.   7.  The localization dispute didn’t, of course, end with Gall and Flourens. For a useful, brief survey of its relatively modem history (since Wernicke), see Eggert (1977). It is of some interest—in passing—that Wernicke, committed localizationalist though he was in respect of the language mechanisms, held that only “primary functions…can be referred to specific areas. All processes which exceed these primary functions (such as the synthesis of various perceptions into concepts and the complex functions such as thought and consciousness) are dependent upon the fiber bundles connecting different areas of the cortex” (p. 92). Barring the associationism, Wernicke’s picture is not very different from the one that we’ve been developing here.

介绍

Introduction

2020 年本应是自动驾驶汽车的时代。五年前，《卫报》的头条新闻预测“从 2020 年开始，你将永远成为后座司机”（Adams，2015 年）。2016 年，Business Insider向我们保证“到 2020 年，将有 1000 万辆自动驾驶汽车上路”（Business Insider Intelligence，2016 年）。特斯拉汽车公司首席执行官埃隆·马斯克在 2019 年承诺“一年后，我们将拥有超过 100 万辆完全自动驾驶、软件……一切齐全的汽车”（Hawkins，2019 年）。2020 年是几家汽车公司宣布将自动驾驶汽车推向市场的目标（McCormick，2017 年；Kageyama，2015 年）。

The year 2020 was supposed to herald the arrival of self-driving cars. Five years earlier, a headline in The Guardian predicted that “from 2020 you will become a permanent backseat driver” (Adams, 2015). In 2016, Business Insider assured us that “10 million self-driving cars will be on the road by 2020” (Business Insider Intelligence, 2016). Tesla Motors CEO, Elon Musk, promised in 2019 that “a year from now, we’ll have over a million cars with full self-driving, software…everything” (Hawkins, 2019). And 2020 was the target announced by several automobile companies to bring self-driving cars to market (McCormick, 2017; Kageyama, 2015).

尽管有人试图重新定义“全自动驾驶” (Baldwin, 2021)，但这些预测都没有实现。值得引用人工智能专家 Drew McDermott 的话，当对人工智能系统（尤其是自动驾驶汽车）的过度乐观被证明是错误时会发生什么：

Despite attempts to redefine “full self-driving” into existence (Baldwin, 2021), none of these predictions has come true. It’s worth quoting AI expert Drew McDermott on what can happen when over-optimism about AI systems—in particular, self-driving cars—turns out to be wrong:

最值得注意的是，麦克德莫特的警告是在 1985 年发出的，当时，就像今天一样，人工智能领域对机器智能的近期前景充满信心和乐观。麦克德莫特写的是该领域的周期性模式。新的、明显的突破将使人工智能从业者预测快速的进步、成功的商业化和“真正的人工智能”的近期前景。政府和公司将陷入这种热情之中，并将为该领域提供大量研发资金。人工智能的春天将会盛开。当进展停滞时，热情、资金和工作机会就会枯竭。人工智能的冬天将会到来。事实上，在麦克德莫特发出警告大约五年后，一个新的人工智能冬天来临了。

What’s most notable is that McDermott’s warning is from 1985, when, like today, the field of AI was awash with confident optimism about the near future of machine intelligence. McDermott was writing about a cyclical pattern in the field. New, apparent breakthroughs would lead AI practitioners to predict rapid progress, successful commercialization, and the near-term prospects of “true AI.” Governments and companies would get caught up in the enthusiasm, and would shower the field with research and development funding. AI spring would be in bloom. When progress stalled, the enthusiasm, funding, and jobs would dry up. AI winter would arrive. Indeed, about five years after McDermott’s warning, a new AI winter set in.

在本章中，我将探讨人工智能的过度自信和失望这一循环往复的原因。我认为，公众、媒体甚至专家的过度乐观可能源于我们谈论人工智能的方式以及我们对智能本质的直觉中的几种谬误。了解这些谬误及其微妙的影响可能有助于指导创建更强大、更值得信赖甚至可能真正智能的人工智能系统。

In this chapter, I explore the reasons for the repeating cycle of overconfidence followed by disappointment in expectations about AI. I argue that over-optimism among the public, the media, and even experts can arise from several fallacies in how we talk about AI and in our intuitions about the nature of intelligence. Understanding these fallacies and their subtle influences may help guide the creation of more robust, trustworthy, and perhaps actually intelligent AI systems.

春天和冬天

Springs and winters

对人工智能过度自信的预测与该领域本身一样古老。例如，1958 年，《纽约时报》报道了美国海军对弗兰克·罗森布拉特 (Frank Rosenblatt) 的“感知器”（当今深度神经网络的前身）的演示：“海军今天展示了电子计算机的雏形，它希望它能够行走、说话、看、写、自我复制，并意识到自己的存在”（《纽约时报》，1958 年）。这种乐观的看法很快得到了人工智能先驱们的类似宣言，这一次是关于基于逻辑的“符号”人工智能的前景。1960 年，赫伯特·西蒙宣称“二十年内，机器将能够做人类可以做的任何工作”（Simon，1960 年）。次年，克劳德·香农（Claude Shannon）也做出了同样的预测：“我确信，在 10 到 15 年内，实验室中将会出现一些与科幻小说中机器人相差无几的东西”（IEEE 信息理论学会，2016 年）。几年后，马文·明斯基（Marvin Minsky）预测：“在一代人的时间内……创造‘人工智能’的问题将得到实质性解决”（明斯基，1967 年）。

Overconfident predictions about AI are as old as the field itself. In 1958, for example, the New York Times reported on a demonstration by the US Navy of Frank Rosenblatt’s “perceptron” (a rudimentary precursor to today’s deep neural networks): “The Navy revealed the embryo of an electronic computer today that it expects will be able to walk, talk, see, write, reproduce itself, and be conscious of its existence” (The New York Times, 1958). This optimistic take was quickly followed by similar proclamations from AI pioneers, this time about the promise of logic-based “symbolic” AI. In 1960 Herbert Simon declared that “machines will be capable, within twenty years, of doing any work that a man can do” (Simon, 1960). The following year, Claude Shannon echoed this prediction: “I confidently expect that within a matter of 10 or 15 years, something will emerge from the laboratory which is not too far from the robot of science fiction fame” (IEEE Information Theory Society, 2016). And a few years later Marvin Minsky forecast that “within a generation…the problems of creating ‘artificial intelligence’ will be substantially solved” (Minsky, 1967).

这些预测反映出 20 世纪 60 年代和 70 年代初人工智能的乐观春天，但很快迎来了第一个人工智能寒冬。明斯基和帕普特在 1969 年出版的《感知器》（Minsky and Papert，1969）一书中指出，罗森布拉特的感知器能够解决的问题种类非常有限。1973 年，英国的莱特希尔报告（Lighthill，1973）和美国国防部的“美国研究小组”报告（由各自政府委托，旨在评估人工智能在不久的将来的前景）都对人工智能的前景持极为悲观的态度。这导致两国对人工智能的资助急剧减少，热情也随之下降。

The optimistic AI spring of the 1960s and early 1970s, reflected in these predictions, soon gave way to the first AI winter. Minsky and Papert’s 1969 book Perceptrons (Minsky and Papert, 1969) showed that the kinds of problems solvable by Rosenblatt’s perceptrons were very limited. In 1973 the Lighthill Report (Lighthill, 1973) in the UK and the Department of Defense’s “American Study Group” report in the US, commissioned by their respective governments to assess prospects for AI in the near future, were both extremely negative about those prospects. This led to sharp funding decreases and a downturn in enthusiasm for AI in both countries.

从 20 世纪 80 年代初开始，人工智能再次经历了热情的高涨，当时出现了几项新举措：工业领域“专家系统”的兴起（Durkin，1996）；日本对“第五代”项目（Gaines，1984）进行了巨额投资，该项目雄心勃勃，旨在将人工智能能力作为新一代计算系统的核心；美国随之发起“战略计算计划”（Stefik，1985），为通用人工智能的发展提供了大量资金；以及在神经网络方面进行的一系列新努力（McClelland 等，1986a、1986b），为该领域带来了新的希望。

AI once again experienced an upturn in enthusiasm starting in the early 1980s with several new initiatives: the rise of “expert systems” in industry (Durkin, 1996); Japan’s huge investment in its “Fifth Generation” project (Gaines, 1984), which aimed for ambitious AI abilities as the core of a new generation of computing systems; the US’s responding “Strategic Computing Initiative” (Stefik, 1985), which provided large funding for progress into general AI; as well as a new set of efforts on neural networks (McClelland et al., 1986a,b), which generated new hopes for the field.

到 20 世纪 80 年代后期，这些乐观的希望全部破灭了；同样，这些技术都没有实现所做出的崇高承诺。专家系统依靠人类制定规则来捕捉特定领域的专家知识，结果却很脆弱——也就是说，在面对新情况时，它们往往无法概括或适应。问题在于，编写规则的人类专家实际上依赖于潜意识知识——我们可以称之为“常识”——而这并不是系统编程的一部分。第五代项目和战略计算计划所追求的人工智能方法也遇到了类似的脆弱性和缺乏通用性的问题。20 世纪 80 年代和 90 年代的神经网络方法同样在相对简单的例子上效果很好，但缺乏扩展到复杂问题的能力。事实上，20 世纪 80 年代末标志着新一轮人工智能寒冬的开始，该领域的声誉受损。 1990 年我获得博士学位时，有人建议我不要将“人工智能”一词用于求职申请中。

By the latter part of the 1980s, these optimistic hopes had all been dashed; again, none of these technologies had achieved the lofty promises that had been made. Expert systems, which rely on humans to create rules that capture expert knowledge of a particular domain, turned out to be brittle—that is, often unable to generalize or adapt when faced with new situations. The problem was that the human experts writing the rules actually rely on subconscious knowledge—what we might call “common sense”—that was not part of the system’s programming. The AI approaches pursued under the Fifth Generation project and Strategic Computing Initiative ran into similar problems of brittleness and lack of generality. The neural-network approaches of the 1980s and 1990s likewise worked well on relatively simple examples but lacked the ability to scale up to complex problems. Indeed, the late 1980s marked the beginning of a new AI winter, and the field’s reputation suffered. When I received my PhD in 1990, I was advised not to use the term “Artificial Intelligence” on my job applications.

在 1956 年启动该领域的达特茅斯夏季研讨会 50 周年纪念活动上，人工智能先驱、最初创造“人工智能”一词的约翰·麦卡锡 (John McCarthy) 简明扼要地解释了这一问题：“人工智能比我们想象的要难”(Moewes and Nürnberger，2013)。

At the 50th anniversary commemoration of the 1956 Dartmouth Summer Workshop that launched the field, AI pioneer John McCarthy, who had originally coined the term “Artificial Intelligence,” explained the issue succinctly: “AI was harder than we thought” (Moewes and Nürnberger, 2013).

20 世纪 90 年代和 21 世纪见证了机器学习的迅猛发展：从数据中创建预测模型的算法。机器学习算法通常受到统计学而非神经科学或心理学的启发，旨在执行特定任务而非捕获一般智能。机器学习从业者通常很快就能将他们的学科与当时名誉扫地的人工智能领域区分开来。

The 1990s and 2000s saw the meteoric rise of machine learning: algorithms that create predictive models from data. Machine learning algorithms were typically inspired by statistics rather than by neuroscience or psychology, and were aimed at performing specific tasks rather than capturing general intelligence. Machine-learning practitioners were often quick to differentiate their discipline from the then-discredited field of AI.

然而，在 2010 年左右，深度学习（一种受大脑启发的多层神经网络，通过数据进行训练）从落后地位脱颖而出，成为机器学习领域的超级明星。深度神经网络自 20 世纪 70 年代就已出现，但直到最近，由于从网络上抓取的大量数据集、快速并行计算芯片以及训练方法的创新，这些方法才得以扩展，足以解决大量以前未解决的人工智能难题。深度神经网络是我们在过去十年中看到的所有重大人工智能进步的动力，包括语音识别、机器翻译、聊天机器人、图像识别、游戏和蛋白质折叠等。

However, around 2010, deep learning—in which brain-inspired multilayered neural networks are trained from data—emerged from its backwater position and rose to superstar status in machine learning. Deep neural networks had been around since the 1970s, but only recently, due to huge datasets scraped from the Web, fast parallel computing chips, and innovations in training methods, could these methods scale up enough to address a large number of previously unsolved AI challenges. Deep neural networks are what power all of the major AI advances we’ve seen in the past decade, including speech recognition, machine translation, chat bots, image recognition, game playing, and protein folding, among others.

突然间“人工智能”一词开始随处出现，人们对被称为“通用人工智能”、“真正人工智能”或“人类水平人工智能”的人工智能的前景突然燃起新一轮的乐观情绪。

Suddenly the term “AI” started to appear everywhere, and there was all at once a new round of optimism about the prospects of what has been variously called “general,” “true,” or “human-level” AI.

在 2016 年和 2018 年对人工智能研究人员进行的调查中，受访者的预测中值为 50%，即到 2040-2060 年，人类级别的人工智能将诞生，尽管人们对较早和较晚的估计意见不一（Müller 和 Bostrom，2016 年；Grace 等人，2018 年）。甚至一些最著名的人工智能专家和企业家也持相同观点。一本广泛使用的人工智能教科书的合著者 Stuart Russell 预测，“超级智能人工智能”将“很可能在我孩子的有生之年出现”（Russell，2019b 年），而人工智能公司 OpenAI 的首席执行官 Sam Altman 预测，几十年内，计算机程序“将做几乎所有事情，包括做出新的科学发现，从而扩展我们对‘一切’的概念”（Altman，2021 年）。谷歌 DeepMind 联合创始人 Shane Legg 在 2008 年预测“人工智能将在 2020 年代中期超越人类水平”（Despres, 2008），Facebook 首席执行官马克·扎克伯格在 2015 年宣称“Facebook 未来五到十年的目标之一，就是让人工智能在所有主要人类感官上都超越人类水平：视觉、听觉、语言、一般认知”（McCracken, 2015）。

In surveys of AI researchers carried out in 2016 and 2018, the median prediction of those surveyed gave a 50 percent chance that human-level AI would be created by 2040–2060, though there was much variance of opinion, both for sooner and later estimates (Müller and Bostrom, 2016; Grace et al., 2018). Even some of the most well-known AI experts and entrepreneurs are in accord. Stuart Russell, co-author of a widely used textbook on AI, predicts that “superintelligent AI” will “probably happen in the lifetime of my children” (Russell, 2019b) and Sam Altman, CEO of the AI company OpenAI, predicts that within decades, computer programs “will do almost everything, including making new scientific discoveries that will expand our concept of ‘everything’” (Altman, 2021). Shane Legg, co-founder of Google DeepMind, predicted in 2008 that “human level AI will be passed in the mid-2020s” (Despres, 2008), and Facebook’s CEO, Mark Zuckerberg, declared in 2015 that “one of [Facebook’s] goals for the next five to ten years is to basically get better than human level at all of the primary human senses: vision, hearing, language, general cognition” (McCracken, 2015).

然而，尽管人们对深度学习抱有乐观态度，但没过多久，深度学习的智能外表就出现了裂痕。事实证明，与过去的所有人工智能系统一样，深度学习系统也存在脆弱性——在面对与训练数据不同的情况时，会出现不可预测的错误。这是因为这类系统容易受到捷径学习的影响（Geirhos 等人，2020 年；Lapuschkin 等人，2019 年）：学习训练数据中的统计关联，使机器能够给出正确的答案，但有时是出于错误的原因。换句话说，这些机器不会学习我们试图教给它们的概念，而是学习捷径来纠正训练集上的答案——而这种捷径不会导致良好的概括。事实上，深度学习系统通常无法学习抽象概念，而这些抽象概念可以让它们将所学的知识迁移到新的情况或任务中（Mitchell，2021 年）。此外，此类系统容易受到“对抗性扰动”（Moosavi-Dezfooli 等，2017）的攻击——对输入进行特殊设计的改变，这些改变人类无法察觉或无关，但会导致系统出错。

However, in spite of all the optimism, it didn’t take long for cracks to appear in deep learning’s façade of intelligence. It turns out that, like all AI systems of the past, deep-learning systems can exhibit brittleness—unpredictable errors when facing situations that differ from the training data. This is because such systems are susceptible to shortcut learning (Geirhos et al., 2020; Lapuschkin et al., 2019): learning statistical associations in the training data that allow the machine to produce correct answers but sometimes for the wrong reasons. In other words, these machines don’t learn the concepts we are trying to teach them, but rather they learn shortcuts to correct answers on the training set—and such shortcuts will not lead to good generalizations. Indeed, deep-learning systems often cannot learn the abstract concepts that would enable them to transfer what they have learned to new situations or tasks (Mitchell, 2021). Moreover, such systems are vulnerable to attack from “adversarial perturbations” (Moosavi-Dezfooli et al., 2017)—specially engineered changes to the input that are either imperceptible or irrelevant to humans but that induce the system to make errors.

尽管对深度神经网络的局限性进行了广泛的研究，但其脆弱性和易受攻击性的根源仍未完全了解。这些网络具有大量参数，是复杂的系统，其决策机制可能非常不透明。然而，从它们非人类般的错误和易受对抗性干扰的影响来看，这些系统实际上并没有理解它们处理的数据，至少不是人类意义上的“理解”。人工智能界仍在争论这种理解是否可以通过增加网络层和更多训练数据来实现，或者是否缺少一些更基本的东西。

Despite extensive research on the limitations of deep neural networks, the sources of their brittleness and vulnerability are still not completely understood. These networks, with their large number of parameters, are complicated systems whose decision-making mechanisms can be quite opaque. However, it seems clear from their non-humanlike errors and vulnerability to adversarial perturbations that these systems are not actually understanding the data they process, at least not in the human sense of “understand.” It’s still a matter of debate in the AI community whether such understanding can be achieved by adding network layers and more training data, or whether something more fundamental is missing.

在撰写本文时（2021 年中），几种新的深度学习方法再次在 AI 社区中引起了相当大的乐观情绪。一些最热门的新领域包括使用自监督（或“预测”）学习的 Transformer 架构（Devlin 等人，2018 年）、元学习（Finn 等人，2017 年）和深度强化学习（Arulkumaran 等人，2017 年）；这些领域都被认为是朝着更通用、更像人类的 AI 迈出的进步。虽然这些和其他新创新已经显示出初步的希望，但 AI 的春天和冬天的周期可能会继续。该领域在相对狭窄的领域不断进步，但通往人类水平的 AI 的道路并不那么明朗。

At the time of writing (mid-2021), several new deep-learning approaches are once again generating considerable optimism in the AI community. Some of the hottest new areas are transformer architectures using self-supervised (or “predictive”) learning (Devlin et al., 2018), meta-learning (Finn et al., 2017), and deep reinforcement learning (Arulkumaran et al., 2017); each of these has been cited as progress towards more general, humanlike AI. While these and other new innovations have shown preliminary promise, the AI cycle of springs and winters is likely to continue. The field continually advances in relatively narrow areas, but the path toward human-level AI is less clear.

在接下来的几节中，我将论证，对人类级人工智能可能出现时间线的预测反映了我们自身的偏见和对智能本质的缺乏理解。特别是，我将描述我们在思考人工智能时犯下的四个错误，这些错误在我看来最为重要。虽然这些错误在人工智能界广为人知，但专家做出的许多假设仍然受这些错误的影响，并让我们对“真正”智能机器的近期前景产生了错误的信心。

In the next sections, I will argue that predictions about the likely timeline of human-level AI reflect our own biases and lack of understanding of the nature of intelligence. In particular, I describe four fallacies in our thinking about AI that seem most central to me. While these fallacies are well-known in the AI community, many assumptions made by experts still fall victim to these fallacies and give us a false sense of confidence about the near-term prospects of “truly” intelligent machines.

谬误 1：狭义智力与一般智力是连续的

Fallacy 1: Narrow intelligence is on a continuum with general intelligence

特定 AI 任务的进展通常被描述为迈向更通用 AI 的“第一步”。会下棋的计算机 Deep Blue“被誉为 AI 革命的第一步”（Aron，2016 年）。IBM 将其 Watson 系统描述为“迈向认知系统的第一步，开启计算的新时代”（High，2013 年）。OpenAI 的 GPT-3 语言生成器被称为“迈向通用智能的一步”（Alexander，2019 年）。

Advances on a specific AI task are often described as “a first step” towards more general AI. The chess-playing computer Deep Blue “was hailed as the first step of an AI revolution” (Aron, 2016). IBM described its Watson system as “a first step into cognitive systems, a new era of computing” (High, 2013). OpenAI’s GPT-3 language generator was called a “step toward general intelligence” (Alexander, 2019).

事实上，如果人们看到一台机器做了一些了不起的事情，尽管是在很小的领域，他们通常会认为这个领域在通用人工智能方面已经取得了很大进展。哲学家休伯特·德雷福斯（使用了由耶霍舒亚·巴尔-希勒尔创造的一个术语）将此称为“第一步谬误”。正如德雷福斯所描述的那样，“第一步谬误是声称，自从我们第一次研究计算机智能以来，我们一直在沿着一个连续体缓慢前进，而这个连续体的末端就是人工智能，因此，我们程序中的任何改进，无论多么微不足道，都算作进步。”德雷福斯引用了他兄弟、工程师斯图尔特·德雷福斯的一个比喻：“这就像声称第一只爬树的猴子正在朝着登月的方向前进”（德雷福斯，2012 年）。

Indeed, if people see a machine do something amazing, albeit in a narrow area, they often assume the field is that much further along toward general AI. The philosopher Hubert Dreyfus (using a term coined by Yehoshua Bar-Hillel) called this a “first-step fallacy.” As Dreyfus characterized it, “The first-step fallacy is the claim that, ever since our first work on computer intelligence we have been inching along a continuum at the end of which is AI so that any improvement in our programs no matter how trivial counts as progress.” Dreyfus quotes an analogy made by his brother, the engineer Stuart Dreyfus: “It was like claiming that the first monkey that climbed a tree was making progress towards landing on the moon” (Dreyfus, 2012).

和他之前和之后的许多人工智能专家一样，德雷福斯指出，在假定的人工智能进步连续体中，“意想不到的障碍”始终是常识问题。我将在最后一节更多地谈论这个常识障碍。

Like many AI experts before and after him, Dreyfus noted that the “unexpected obstacle” in the assumed continuum of AI progress has always been the problem of common sense. I will say more about this barrier of common sense in the last section.

谬误2：容易的事很难，难的事很容易

Fallacy 2: Easy things are hard and hard things are easy

约翰·麦卡锡哀叹“人工智能比我们想象的要难”，而马文·明斯基则解释说，这是因为“简单的事情很难”（明斯基，1986 年）。也就是说，我们人类不费吹灰之力就能做到的事情——观察世界并理解我们所看到的东西、进行对话、在拥挤的人行道上行走而不撞到任何人——对机器来说却是最困难的挑战。相反，让机器做对人类来说非常困难的事情往往更容易；例如，解决复杂的数学问题、掌握国际象棋和围棋等游戏以及在数百种语言之间翻译句子，这些对机器来说都相对容易。这是一种所谓的“莫拉维克悖论”，以机器人专家汉斯·莫拉维克的名字命名，他写道：“让计算机在智力测试或下跳棋时表现出成人水平的表现相对容易，而让它们在感知和移动方面拥有一岁儿童的技能则很难或不可能”（莫拉维克，1988 年）。

While John McCarthy lamented that “AI was harder than we thought,” Marvin Minsky explained that this is because “easy things are hard” (Minsky, 1986). That is, the things that we humans do without much thought—looking out in the world and making sense of what we see, carrying on a conversation, walking down a crowded sidewalk without bumping into anyone—turn out to be the hardest challenges for machines. Conversely, it’s often easier to get machines to do things that are very hard for humans; for example, solving complex mathematical problems, mastering games like chess and Go, and translating sentences between hundreds of languages have all turned out to be relatively easier for machines. This is a form of what’s been called “Moravec’s paradox,” named after roboticist Hans Moravec, who wrote, “It is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility” (Moravec, 1988).

这种谬论自人工智能诞生之初就一直影响着人们对人工智能的思考。人工智能先驱赫伯特·西蒙宣称：“认知中所有有趣的事情都发生在 100 毫秒以上——也就是你认出母亲的时间”（霍夫施塔特，1985 年）。西蒙说，要理解认知，我们不必担心无意识的感知过程。这一假设反映在大多数符号人工智能传统中，这些传统侧重于对已经被感知的输入进行推理的过程。

This fallacy has influenced thinking about AI since the dawn of the field. AI pioneer Herbert Simon proclaimed, “Everything of interest in cognition happens above the 100-millisecond level—the time it takes you to recognize your mother” (Hofstadter, 1985). Simon is saying that, to understand cognition, we don’t have to worry about unconscious perceptual processes. This assumption is reflected in most of the symbolic AI tradition, which focuses on the process of reasoning about input that has already been perceived.

在过去的几十年里，符号人工智能方法在研究界已经失宠，而研究界主要被深度学习所主导，而深度学习确实解决了感知问题。然而，这种谬误背后的假设仍然出现在最近关于人工智能的说法中。例如，在 2016 年的一篇文章中，深度学习先驱 Andrew Ng 被引用来呼应 Simon 的假设，大大低估了无意识感知和思维的复杂性：“如果一个普通人可以在不到一秒钟的时间内完成一项心理任务，那么我们现在或在不久的将来可能就可以使用人工智能将其自动化”（Ng，2016 年）。

In the last decades, symbolic AI approaches have lost favor in the research community, which has largely been dominated by deep learning, which does address perception. However, the assumptions underlying this fallacy still appear in recent claims about AI. For example, in a 2016 article, deep-learning pioneer Andrew Ng was quoted echoing Simon’s assumptions, vastly underestimating the complexity of unconscious perception and thought: “If a typical person can do a mental task with less than one second of thought, we can probably automate it using AI either now or in the near future” (Ng, 2016).

更微妙的是，谷歌 DeepMind 的研究人员在谈论 AlphaGo 的胜利时，将围棋游戏描述为“最具挑战性的领域”之一（Silver 等人，2017 年）。对谁来说具有挑战性？对人类来说，也许如此，但正如心理学家 Gary Marcus 指出的那样，有些领域，包括游戏，虽然对人类来说很容易，但对人工智能系统来说却比围棋更具挑战性。一个例子是猜字谜游戏，它“需要表演技巧、语言技巧和心理理论”（Marcus，2018b），这些能力远远超出了当今人工智能所能完成的任何能力。

More subtly, researchers at Google DeepMind, in talking about AlphaGo’s triumph, described the game of Go as one of “the most challenging of domains” (Silver et al., 2017). Challenging for whom? For humans, perhaps, but as psychologist Gary Marcus pointed out, there are domains, including games, that, while easy for humans, are much more challenging than Go for AI systems. One example is charades, which “requires acting skills, linguistic skills, and theory of mind” (Marcus, 2018b), abilities that are far beyond anything AI can accomplish today.

人工智能比我们想象的要难，因为我们在很大程度上没有意识到我们自己的思维过程的复杂性。汉斯·莫拉维克这样解释他的悖论：“在人类大脑中高度进化的大型感觉和运动部分中，编码了十亿年关于世界本质以及如何在其中生存的经验。深思熟虑的过程我认为，我们称之为推理的思维是人类思想中最薄弱的一层，它之所以有效，只是因为它得到了这种更古老、更强大、但通常是无意识的感觉运动知识的支持。我们都是感知和运动领域的天才，我们是如此优秀，以至于我们让困难的事情看起来很容易”（Moravec，1988 年）。或者更简洁地说，马文·明斯基指出，“一般来说，我们最不了解我们的大脑最擅长什么”（明斯基，1980 年）。

AI is harder than we think because we are largely unconscious of the complexity of our own thought processes. Hans Moravec explains his paradox this way: “Encoded in the large, highly evolved sensory and motor portions of the human brain is a billion years of experience about the nature of the world and how to survive in it. The deliberate process we call reasoning is, I believe, the thinnest veneer of human thought, effective only because it is supported by this much older and much more powerful, though usually unconscious, sensorimotor knowledge. We are all prodigious Olympians in perceptual and motor areas, so good that we make the difficult look easy” (Moravec, 1988). Or more succinctly, Marvin Minsky notes, “In general, we’re least aware of what our minds do best” (Minsky, 1980).

谬误三：一厢情愿的助记符的诱惑

Fallacy 3: The lure of wishful mnemonics

“一厢情愿的助记符”这一术语是由计算机科学家德鲁·麦克德莫特 (Drew McDermott) 在 1976 年对人工智能的批评中创造的：

The term “wishful mnemonic” was coined in a 1976 critique of AI by computer scientist Drew McDermott:

如今，几十年过去了，人工智能研究充斥着此类一厢情愿的助记符——与人类智能相关的术语，用于描述人工智能程序的行为和评估。神经网络大致受到大脑的启发，但又有很大不同。机器学习或深度学习方法与人类（或非人类动物）的学习并不十分相似。事实上，如果机器已经学会了人类意义上的学习，我们会期望它能够在不同的环境中运用它所学到的东西。然而，事实证明情况往往并非如此。在机器学习中，有一个完整的子领域称为迁移学习，它专注于如何让机器将所学知识迁移到新情况中这一尚未解决的问题，这是人类学习的基础能力。

Now, many decades later, work on AI is replete with such wishful mnemonics—terms associated with human intelligence that are used to describe the behavior and evaluation of AI programs. Neural networks are loosely inspired by the brain, but with vast differences. Machine learning or deep learning methods do not really resemble learning in humans (or in non-human animals). Indeed, if a machine has learned something in the human sense of learn, we would expect that it would be able to use what it has learned in different contexts. However, it turns out that this is often not the case. In machine learning, there is an entire subfield called transfer learning that focuses on the still-open problem of how to enable machines to transfer what they have learned to new situations, an ability that is fundamental to human learning.

事实上，我们谈论机器能力的方式影响了我们对这些能力究竟有多普遍的认识。IBM 的一位高管无意中在现实世界中证实了麦克德莫特的警告，他宣称“沃森可以在几秒钟内阅读世界上所有的医疗保健文本”（Gustin，2011）。DeepMind 联合创始人 Demis Hassabis 告诉我们，“AlphaGo 的目标是击败最优秀的人类玩家，而不仅仅是模仿他们”（Ji-hye，2016）。AlphaGo 的首席研究员 David Silver 如此描述该程序的其中一场比赛：“我们总是可以在比赛中询问 AlphaGo 它认为自己表现如何。……直到比赛快结束时，AlphaGo 才认为自己会赢”（Shead，2017）。（以上引文中的重点是我加的。）

Indeed, the way we talk about machine abilities influences our conceptions of how general those abilities really are. Unintentionally providing real-world illustrations of McDermott’s warning, one of IBM’s top executives proclaimed that “Watson can read all of the health-care texts in the world in seconds” (Gustin, 2011). DeepMind co-founder Demis Hassabis tells us that “AlphaGo’s goal is to beat the best human players, not just mimic them” (Ji-hye, 2016). And AlphaGo’s lead research David Silver described one of the program’s matches thus: “We can always ask AlphaGo how well it thinks it’s doing during the game.…It was only towards the end of the game that AlphaGo thought it would win” (Shead, 2017). (Emphasis is mine in the quotations above.)

有人可能会说，这种拟人化术语只是一种简写：IBM 科学家知道 Watson 无法像人类一样阅读或理解；DeepMind 科学家知道 AlphaGo 不像人类那样有目标或想法，也没有像人类一样的“游戏”或“获胜”概念。然而，这种简写可能会误导试图理解这些结果的公众（以及报道这些结果的媒体），甚至可能在不知不觉中影响人工智能专家对其系统的看法，以及这些系统与人类智能的相似程度。

One could argue that such anthropomorphic terms are simply shorthand: IBM scientists know that Watson doesn’t read or understand in the way humans do; DeepMind scientists know that AlphaGo has no goals or thoughts in the way humans do and no humanlike conceptions of a “game” or of “winning.” However, such shorthand can be misleading to the public trying to understand these results (and to the media reporting on them) and can also unconsciously shape the way even AI experts think about their systems and how closely these systems resemble human intelligence.

McDermott 的“一厢情愿的助记符”指的是我们用来描述人工智能程序的术语，但研究界也会使用一厢情愿的助记符来根据我们希望它们测试的技能来命名人工智能评估基准。例如，以下是人工智能子领域“自然语言处理”（NLP）中一些被广泛引用的当前基准：“斯坦福问答数据集”（Rajpurkar 等人，2016 年）、“RACE 阅读理解数据集”（Lai 等人，2017 年）和“通用语言理解评估”（Wang 等人，2019 年）。在所有这些基准测试中，最佳机器的性能已经超过了人类（通常是 Amazon Mechanical Turk 工人）的性能。这导致了这样的头条新闻：“新的人工智能模型在回答问题方面超越了人类的表现”（Costenaro，2018 年）；“计算机的阅读能力越来越强”（Pham，2018 年）；以及“微软的 AI 模型在自然语言理解方面表现优于人类”（Jawad，2021 年）。考虑到这些基准评估的名称，人们得出这样的结论也就不足为奇了。问题是，这些基准实际上并没有衡量问答、阅读理解或自然语言理解的一般能力。基准测试只测试了这些能力的非常有限的版本；此外，正如我上面所描述的，其中许多基准允许机器学习捷径——机器可以利用统计相关性在测试中取得高分，而无需学习实际被测试的技能（McCoy 等人，2019 年；Linzen，2020 年）。虽然机器可以在这些特定的基准测试中胜过人类，但 AI 系统还远远达不到我们与基准名称联系起来的更一般的人类能力。

McDermott’s “wishful mnemonics” referred to terms we use to describe AI programs, but the research community also uses wishful mnemonics in naming AI evaluation benchmarks after the skills we hope they test. For example, here are some of the most widely cited current benchmarks in the subarea of AI called “natural-language processing” (NLP): the “Stanford Question Answering Dataset” (Rajpurkar et al., 2016), the “RACE Reading Comprehension Dataset” (Lai et al., 2017), and the “General Language Understanding Evaluation” (Wang et al., 2019). In all of these benchmarks, the performance of the best machines has already exceeded that measured for humans (typically Amazon Mechanical Turk workers). This has led to headlines such as “New AI model exceeds human performance at question answering” (Costenaro, 2018); “Computers are getting better than humans at reading” (Pham, 2018); and “Microsoft’s AI model has outperformed humans in natural-language understanding” (Jawad, 2021). Given the names of these benchmark evaluations, it’s not surprising that people would draw such conclusions. The problem is, these benchmarks don’t actually measure general abilities for question-answering, reading comprehension, or natural-language understanding. The benchmarks test only very limited versions of these abilities; moreover, many of these benchmarks allow machines to learn shortcuts, as I described above—statistical correlations that machines can exploit to achieve high performance on the test without learning the actual skill being tested (McCoy et al., 2019; Linzen, 2020). While machines can outperform humans on these particular benchmarks, AI systems are still far from matching the more general human abilities we associate with the benchmarks’ names.

谬误四：智力全在于大脑

Fallacy 4: Intelligence is all in the brain

智力是一种不同于身体的东西，无论是作为非物质物质还是仅依赖于大脑，这种观点在哲学和认知科学中有着悠久的历史。

The idea that intelligence is something distinct from the body, whether as a non-physical substance or as dependent only on the brain, has a long history in philosophy and cognitive science.

所谓的“心智信息处理模型”兴起于二十世纪中叶的心理学中。该模型将心智视为一种计算机，可以输入、存储、处理和输出信息。除了输入（感知）和输出（行为）阶段外，身体不会发挥太大作用。根据这种观点，认知完全发生在大脑中，理论上可以与身体的其他部分分离。这种观点的一个极端推论是，未来我们将能够将我们的大脑（以及我们的认知和意识）“上传”到计算机（Woollaston，2013 年）。

The so-called “information-processing model of mind” arose in psychology in the mid-twentieth century. This model views the mind as a kind of computer, which inputs, stores, processes, and outputs information. The body does not play much of a role except in the input (perception) and output (behavior) stages. Under this view, cognition takes place wholly in the brain and is, in theory, separable from the rest of the body. An extreme corollary of this view is that, in the future, we will be able to “upload” our brains—and thus our cognition and consciousness—to computers (Woollaston, 2013).

几乎所有关于人工智能的研究都隐含着这样一个假设：智能原则上可以“脱离肉体”。早期人工智能研究中最有影响力的思想之一是纽厄尔和西蒙的“物理符号系统假说”（PSSH），该假说指出：“物理符号系统具有实现一般智能行为的必要和充分手段”（Newell and Simon，1976 年）。“物理符号系统”一词指的是与数字计算机非常相似的东西。PSSH 假设，数字计算机无需结合任何非符号的大脑或身体过程即可实现一般智能。（有关符号过程与亚符号过程的深入讨论，请参阅霍夫施塔特（1985 年）的《从布尔梦中醒来》。）

The assumption that intelligence can in principle be “disembodied” is implicit in almost all work on AI throughout its history. One of the most influential ideas in early AI research was Newell and Simon’s “Physical Symbol System Hypothesis” (PSSH), which stated: “A physical symbol system has the necessary and sufficient means for general intelligent action” (Newell and Simon, 1976). The term “physical symbol system” refers to something much like a digital computer. The PSSH posits that general intelligence can be achieved in digital computers without incorporating any non-symbolic processes of brain or body. (For an insightful discussion of symbolic versus subsymbolic processes, see Hofstadter (1985)’s “Waking Up from the Boolean Dream.”)

纽厄尔和西蒙的 PSSH 是人工智能符号方法的创始原则，在 20 世纪 90 年代和 21 世纪统计和神经启发式机器学习兴起之前，符号方法一直主导着人工智能领域。然而，这些非符号方法也不认为身体与智能相关。相反，从 20 世纪 80 年代的联结主义到当今的深度神经网络，神经启发式方法通常认为智能完全源于大脑结构和动态。当今的深度神经网络类似于众所周知的缸中之脑：被动地从世界获取数据并输出行为指令，而不主动与任何身体进行互动。当然，机器人和自动驾驶汽车的不同之处在于它们在世界上有物理存在，但迄今为止，它们之间的物理互动类型以及对“智能”的反馈都相当有限。

Newell and Simon’s PSSH was a founding principle of the symbolic approach to AI, which dominated the field until the rise of statistical and neurally inspired machine learning in the 1990s and 2000s. However, these non-symbolic approaches also do not view the body as relevant to intelligence. Instead, neurally inspired approaches from 1980s connectionism to today’s deep neural networks generally assume that intelligence arises solely from brain structures and dynamics. Today’s deep neural networks are akin to the proverbial brain-in-a-vat: passively taking in data from the world and outputting instructions for behavior without actively interacting in the world with any kind of body. Of course, robots and autonomous vehicles are different in that they have a physical presence in the world, but to date the kinds of physical interactions they have, and the feedback to their “intelligence,” is quite limited.

智能完全存在于大脑中的假设引发了人们的猜测：要实现人类级别的人工智能，我们只需将机器规模扩大到与大脑的“计算能力”相匹配，然后为这种与大脑相匹配的“硬件”开发适当的“软件”。例如，一位哲学家撰写了一份文献报告，结论是：“我认为 10 ¹⁵ FLOP/s 足以像人脑一样执行任务（只要有合适的软件，这可能很难创建）”（Carlsmith，2020 年）。不需要任何人！

The assumption that intelligence is all in the brain has led to speculation that, to achieve human-level AI, we simply need to scale up machines to match the brain’s “computing capacity” and then develop the appropriate “software” for this brain-matching “hardware.” For example, one philosopher wrote a report on the literature that concluded, “I think it more likely than not that 10¹⁵ FLOP/s is enough to perform tasks as well as the human brain (given the right software, which may be very hard to create)” (Carlsmith, 2020). No body needed!

顶级人工智能研究人员也赞同这一观点，即扩大硬件规模以匹配大脑将实现人类级别的人工智能。例如，深度学习先驱杰弗里·辛顿 (Geoffrey Hinton) 预测，“要以人类的水平理解 [文档]，我们可能需要人类级别的资源，而我们的大脑中有数万亿个连接。……但迄今为止我们建立的最大网络只有数十亿个连接。所以我们还差了几个数量级，但我相信硬件人员会解决这个问题”（Patterson 和 Gibson，2017 年）。其他人预测，“硬件修复”——最终实现人类级别人工智能的速度和内存容量——将以量子计算机的形式出现（Musser，2018 年）。

Top AI researchers have echoed the idea that scaling up hardware to match the brain will enable human-level artificial intelligence. For example, deep-learning pioneer Geoffrey Hinton predicted, “To understand [documents] at a human level, we’re probably going to need human-level resources and we have trillions of connections [in our brains].…But the biggest networks we have built so far only have billions of connections. So we’re a few orders of magnitude off, but I’m sure the hardware people will fix that” (Patterson and Gibson, 2017). Others have predicted that the “hardware fix”—the speed and memory capacity to finally enable human-level AI—will come in the form of quantum computers (Musser, 2018).

然而，越来越多的研究人员开始质疑“一切都在大脑中”的信息处理模型在理解智能和创造人工智能方面的基础。计算机科学家罗德·布鲁克斯认为，“我们之所以长期被困在死胡同里，是因为摩尔定律一直在推动我们前进，我们一直在想，‘哦，我们在进步，我们在进步，我们在进步。’但也许我们并没有进步。”（布鲁克斯，2019 年）。事实上，几十年来，许多认知科学家一直主张身体在所有认知活动中的核心地位。这些观点的一位著名支持者、心理学家马克·约翰逊 (Mark Johnson) 写道，一项关于具身认知的研究计划在 20 世纪 70 年代中期开始兴起，它“开始提供越来越多的证据，证明我们的大脑和身体在我们经历、思考和做的一切事情中都发挥着核心作用。”（约翰逊，2017 年）。心理学家丽贝卡·芬奇-基弗 (Rebecca Fincher-Kiefer) 这样描述具身认知范式：“具身认知意味着概念知识的表征依赖于身体：它是多模态的……，而不是非模态的、符号的或抽象的。这一理论表明，我们的思想是有根基的，或与感知、行动和情感密不可分，我们的大脑和身体共同作用产生认知”(Fincher-Kiefer，2019)。

However, a growing cadre of researchers is questioning the basis of the “all in the brain” information-processing model for understanding intelligence and for creating AI. Writing about what he calls “the cul-de-sac of the computational metaphor,” computer scientist Rod Brooks argues, “The reason for why we got stuck in this cul-de-sac for so long was because Moore’s law just kept feeding us, and we kept thinking, ‘Oh, we’re making progress, we’re making progress, we’re making progress.’ But maybe we haven’t been” (Brooks, 2019). In fact, a number of cognitive scientists have argued for decades for the centrality of the body in all cognitive activities. One prominent proponent of these ideas, the psychologist Mark Johnson, writes of a research program on embodied cognition, gaining steam in the mid-1970s, that “began to provide converging evidence for the central role of our brains and bodies in everything we experience, think, and do” (Johnson, 2017). Psychologist Rebecca Fincher-Kiefer characterizes the embodied cognition paradigm this way: “Embodied cognition means that the representation of conceptual knowledge is dependent on the body: it is multimodal…, not amodal, symbolic, or abstract. This theory suggests that our thoughts are grounded, or inextricably associated with, perception, action, and emotion, and that our brain and body work together to have cognition” (Fincher-Kiefer, 2019).

具身认知的证据来自多种学科。例如，神经科学的研究表明，控制认知的神经结构与控制感觉和运动系统的神经结构有着密切的联系，抽象思维利用了基于身体的神经“地图”（Epstein 等人，2017 年）。正如神经科学家唐·塔克 (Don Tucker) 所说，“大脑没有用于非具身认知的部分”（Tucker，2007 年）。认知心理学和语言学的结果表明，我们的许多（如果不是全部的话）抽象概念都植根于物理的、基于身体的内部模型（Barsalou 和 Wiemer-Hastings，2005 年），这部分是由日常语言中基于身体的隐喻系统揭示的（Lakoff 和 Johnson，2008 年）。

The evidence for embodied cognition comes from a diverse set of disciplines. Research in neuroscience suggests, for example, that the neural structures controlling cognition are richly linked to those controlling sensory and motor systems, and that abstract thinking exploits body-based neural “maps” (Epstein et al., 2017). As neuroscientist Don Tucker noted, “There are no brain parts for disembodied cognition” (Tucker, 2007). Results from cognitive psychology and linguistics indicate that many, if not all, of our abstract concepts are grounded in physical, body-based internal models (Barsalou and Wiemer-Hastings, 2005), revealed in part by the systems of physically based metaphors found in everyday language (Lakoff and Johnson, 2008).

其他几个学科，如发展心理学，也为具身认知提供了证据。然而，人工智能研究大多忽略了这些结果，尽管有一小部分研究人员在“具身人工智能”、“发展机器人”和“扎实语言理解”等子领域探索这些想法。

Several other disciplines, such as developmental psychology, add to evidence for embodied cognition. However, research in AI has mostly ignored these results, though there is a small group of researchers exploring these ideas in subareas known as “embodied AI,” “developmental robotics,” and “grounded language understanding,” among others.

与具身认知理论相关的是，伴随我们深度社交生活的情绪和“非理性”偏见（通常被认为与智力无关，或妨碍理性）实际上是实现智力的关键。人工智能通常被认为是一种“纯智能”，一种独立于情绪、非理性和身体限制（如进食和睡眠需求）的智能。这种对纯理性智能可能性的假设可能导致我们对未来“超级智能”机器将面临的风险做出耸人听闻的预测。

Related to the theory of embodied cognition is the idea that the emotions and the “irrational” biases that go along with our deeply social lives—typically thought of as separate from intelligence, or as getting in the way of rationality—are actually key to what makes intelligence possible. AI is often thought of as aiming at a kind of “pure intelligence,” one that is independent of emotions, irrationality, and constraints of the body such as the need to eat and sleep. This assumption of the possibility of a purely rational intelligence can lead to lurid predictions about the risks we will face from future “superintelligent” machines.

例如，哲学家尼克·博斯特罗姆 (Nick Bostrom) 断言，系统的智能和其目标是正交的；他认为“任何级别的智能都可以与任何最终目标相结合”(Bostrom, 2014, 105)。例如，博斯特罗姆设想了一个假想的超级智能 AI 系统，其唯一目标是生产回形针；这个假想系统的超级智能使人们能够发明出生产回形针的巧妙方法，并在生产过程中耗尽地球的所有资源。

For example, the philosopher Nick Bostrom asserts that a system’s intelligence and its goals are orthogonal; he argues that “any level of intelligence could be combined with any final goal” (Bostrom, 2014, 105). As an example, Bostrom imagines a hypothetical superintelligent AI system whose sole objective is to produce paperclips; this imaginary system’s superintelligence enables the invention of ingenious ways to produce paperclips, and uses up all of the Earth’s resources in doing so.

人工智能研究员 Stuart Russell 同意 Bostrom 关于智能和目标正交性的观点。“很容易想象，通用智能系统可以或多或少地被赋予任何目标，包括最大化回形针的数量或已知的圆周率位数”（Russell，2019b，167）。Russell 担心使用这种超级智能来解决人类问题可能产生的结果：“如果一个超级智能气候控制系统的任务是将二氧化碳浓度恢复到工业化前的水平，而它认为解决方案是将人类人口减少到零，那会怎样？……如果我们将错误的目标输入机器，而它比我们更聪明，我们就会失败”（Russell，2019a）。

AI researcher Stuart Russell concurs with Bostrom on the orthogonality of intelligence and goals. “It is easy to imagine that a general-purpose intelligent system could be given more or less any objective to pursue, including maximizing the number of paper clips or the number of known digits of pi” (Russell, 2019b, 167). Russell worries about the possible outcomes of employing such a superintelligence to solve humanity’s problems: “What if a superintelligent climate control system, given the job of restoring carbon dioxide concentrations to preindustrial levels, believes the solution is to reduce the human population to zero?…If we insert the wrong objective into the machine and it is more intelligent than us, we lose” (Russell, 2019a).

博斯特罗姆和拉塞尔提出的思想实验似乎假设人工智能系统可以在没有任何基本人类常识的情况下实现“超级智能”，同时无缝保留计算机的速度、精度和可编程性。但这些关于超人人工智能的推测受到对智能本质的错误直觉的困扰。我们的心理学或神经科学知识中没有任何内容支持“纯粹理性”可以与塑造我们的认知和目标的情感和文化偏见分离的可能性。相反，我们从具身认知研究中了解到，人类智能似乎是一个高度集成的系统，具有紧密相连的属性，包括情感、欲望、强烈的自我意识和自主性，以及对世界的常识性理解。这些属性是否可以分离尚不清楚。

The thought experiments proposed by Bostrom and Russell seem to assume that an AI system could be “superintelligent” without any basic humanlike common sense, yet while seamlessly preserving the speed, precision, and programmability of a computer. But these speculations about superhuman AI are plagued by flawed intuitions about the nature of intelligence. Nothing in our knowledge of psychology or neuroscience supports the possibility that “pure rationality” is separable from the emotions and cultural biases that shape our cognition and our objectives. Instead, what we’ve learned from research in embodied cognition is that human intelligence seems to be a strongly integrated system with closely interconnected attributes, including emotions, desires, a strong sense of selfhood and autonomy, and a commonsense understanding of the world. It’s not at all clear that these attributes can be separated.

结论

Conclusions

我所描述的四种谬误揭示了我们对人工智能现状的概念化存在缺陷，以及我们对智能本质的有限直觉。我认为这些谬误至少在一定程度上解释了为什么在机器中捕捉人类智能总是比我们想象的要难。

The four fallacies I have described reveal flaws in our conceptualizations of the current state of AI and our limited intuitions about the nature of intelligence. I have argued that these fallacies are at least in part why capturing humanlike intelligence in machines always turns out to be harder than we think.

这些谬论给人工智能研究人员提出了几个问题。我们如何评估“通用”或“人类水平”人工智能的实际进展？我们如何评估人工智能在特定领域相对于人类的难度？我们应该如何描述人工智能系统的实际能力，而不是用一厢情愿的助记符欺骗自己和他人？人类认知的各个维度（包括认知偏见、情感、目标和体现）在多大程度上可以解开？我们如何改善对智能的直觉？

These fallacies raise several questions for AI researchers. How can we assess actual progress toward “general” or “human-level” AI? How can we assess the difficulty of a particular domain for AI as compared with humans? How should we describe the actual abilities of AI systems without fooling ourselves and others with wishful mnemonics? To what extent can the various dimensions of human cognition (including cognitive biases, emotions, objectives, and embodiment) be disentangled? How can we improve our intuitions about what intelligence is?

这些问题仍未得到解决。显然，为了更有效地推动和评估人工智能的发展，我们需要开发更好的词汇来谈论机器能做什么。更广泛地说，我们需要更好地科学地理解智能，因为它体现在自然界的不同系统中。这将要求人工智能研究人员更深入地参与研究智能的其他科学学科。

These questions remain open. It’s clear that to make and assess progress in AI more effectively, we will need to develop a better vocabulary for talking about what machines can do. And more generally, we will need a better scientific understanding of intelligence as it manifests in different systems in nature. This will require AI researchers to engage more deeply with other scientific disciplines that study intelligence.

常识是智能的一个方面，最近，这一概念推动了人工智能研究人员与来自其他多个学科的认知科学家之间的合作，尤其是认知发展 (例如，参见 Turek, 2018)。在人工智能的历史上，人们曾多次尝试赋予机器类似人类的常识¹，从 John McCarthy (McCarthy, 1986) 和 Douglas Lenat (Lenat et al., 1990) 的基于逻辑的方法，到当今基于深度学习的方法 (例如，Zellers et al., 2019)。人工智能研究员 Oren Etzioni 将“常识”称为“人工智能的暗物质”，并指出，“它有点难以言喻，但你会看到它对一切事物的影响”(Knight, 2018)。这一术语已成为当今最先进的人工智能系统所缺失内容的某种概括（Davis 和 Marcus，2015；Levesque，2017）。虽然常识包括我们人类对世界的大量知识，但它还需要能够利用这些知识来识别和预测我们遇到的情况，并指导我们在这些情况下的行动。赋予机器常识将需要向它们灌输人类婴儿所拥有的关于空间、时间、因果关系以及无生命物体和其他生命体的性质的最基本的“核心”知识，也许是与生俱来的知识（Spelke 和 Kinzler，2007），从细节抽象到一般概念的能力，以及从先前经验中进行类比的能力。目前还没有人知道如何在机器中捕捉这些知识或能力。这是人工智能研究的当前前沿，一个令人鼓舞的前进方向是利用已知的关于幼儿这些能力的发展情况。有趣的是，这正是艾伦·图灵在 1950 年发表的引入图灵测试的论文中推荐的方法。图灵问道：“与其尝试编写一个模拟成人思维的程序，为什么不尝试编写一个模拟儿童思维的程序呢？”（图灵，1950 年）

The notion of common sense is one aspect of intelligence that has recently been driving collaborations between AI researchers and cognitive scientists from several other disciplines, particularly cognitive development (e.g., see Turek, 2018). There have been many attempts in the history of AI to give humanlike common sense to machines¹ , ranging from the logic-based approaches of John McCarthy (McCarthy, 1986) and Douglas Lenat (Lenat et al., 1990) to today’s deep-learning-based approaches (e.g., Zellers et al., 2019). “Common sense” is what AI researcher Oren Etzioni called “the dark matter of artificial intelligence,” noting, “It’s a little bit ineffable, but you see its effects on everything” (Knight, 2018). The term has become a kind of umbrella for what’s missing from today’s state-of-the-art AI systems (Davis and Marcus, 2015; Levesque, 2017). While common sense includes the vast amount of knowledge we humans have about the world, it also requires being able to use that knowledge to recognize and make predictions about the situations we encounter and to guide our actions in those situations. Giving machines common sense will require imbuing them with the very basic “core,” perhaps innate, knowledge that human infants possess about space, time, causality, and the nature of inanimate objects and other living agents (Spelke and Kinzler, 2007), the ability to abstract from particulars to general concepts, and to make analogies from prior experience. No one yet knows how to capture such knowledge or abilities in machines. This is the current frontier of AI research, and one encouraging way forward is to tap into what’s known about the development of these abilities in young children. Interestingly, this was the approach recommended by Alan Turing in his 1950 paper that introduced the Turing test. Turing asks, “Instead of trying to produce a programme to simulate the adult mind, why not rather try to produce one which simulates the child’s?” (Turing, 1950)

1892 年，心理学家威廉·詹姆斯 (William James) 曾这样评价当时的心理学：“这不是科学，而只是科学的希望”(James, 1892)。这句话完美地概括了当今的人工智能。事实上，一些研究人员将人工智能与中世纪的炼金术进行了类比。1977 年，人工智能研究员特里·维诺格拉德 (Terry Winograd) 写道：“在某些方面，[人工智能] 类似于中世纪的炼金术。我们正处于将不同的物质组合在一起并观察会发生什么的阶段，尚未开发出令人满意的理论……但……炼金术士的实践经验和好奇心提供了丰富的数据，从中可以开发出科学的化学理论”(Winograd, 1977)。四十年后，微软研究院院长埃里克·霍维茨 (Eric Horvitz) 也表示赞同：“现在，我们所做的不是一门科学，而是一种炼金术”(Metz, 2017)。为了理解人工智能真正进步的本质，特别是为什么它比我们想象的更难，我们需要从炼金术转向发展对智能的科学理解。

In 1892, the psychologist William James said of psychology at the time, “This is no science; it is only the hope of a science” (James, 1892). This is a perfect characterization of today’s AI. Indeed, several researchers have made analogies between AI and the medieval practice of alchemy. In 1977, AI researcher Terry Winograd wrote, “In some ways [AI] is akin to medieval alchemy. We are at the stage of pouring together different combinations of substances and seeing what happens, not yet having developed satisfactory theories…but…it was the practical experience and curiosity of the alchemists which provided the wealth of data from which a scientific theory of chemistry could be developed” (Winograd, 1977). Four decades later, Eric Horvitz, director of Microsoft Research, concurred: “Right now, what we are doing is not a science but a kind of alchemy” (Metz, 2017). In order to understand the nature of true progress in AI, and in particular, why it is harder than we think, we need to move from alchemy to developing a scientific understanding of intelligence.

第 10 章注释

Notes for Chapter 10

1.   1.有些人质疑我们为什么需要机器具有像人类一样的认知，但如果我们希望机器在人类世界中与我们一起工作，我们就需要它们拥有与我们自身思维基础相同的关于世界的基本知识。

2.   1.  Some have questioned why we need machines to have humanlike cognition, but if we want machines to work with us in our human world, we will need them to have the same basic knowledge about the world that is the foundation of our own thinking.

第三部分：意向性和理解

Part III: Intentionality and Understanding

机器是否仅仅表现得好像它们与世界保持着联系？或者我们能否确定它们也会思考事物——它们会生成具有内容的状态，而这些内容指的是世界上的物体？人类思维的一个惊人之处在于，我们可以表征世界，既可以表征世界的本来面目，也可以表征世界的可能面目。我们可以看到山峰；我们可以渴望攀登它；即使山峰不在视野范围内，我们也可以想起它；我们可以想象它可能与实际情况有微妙甚至巨大的不同（例如，由不同种类的岩石构成，或拥有不同的动植物）；我们甚至可以想象在水晶喷泉旁的群山，母鸡在那里产下半熟的鸡蛋。即使我们没有意识到，我们也会对同一个物体产生矛盾的想法：一个登山者可能渴望攀登乔戈里峰，但又不想攀登乔戈里峰，尽管乔戈里峰就是乔戈里峰。弗朗茨·布伦塔诺 (Franz Brentano) 在他的《从经验主义角度看的心理学》 (1874/1973)中首次引起了人们对这组现象的关注。它后来被称为意向性。事实上，布伦塔诺将意向性描述为精神的标志：一种只有精神状态才具有的特征。

Do machines merely act as if they are in touch with the world? Or can we be sure that they also have thoughts about things—that they generate states with content, which refer to objects in the world? One of the striking things about human minds is that we can represent the world, both as it is and as it might be. We can see the mountain peak; we can desire to climb it; we can call it to mind even when it’s not in view; we can imagine that it might be subtly or even vastly different than it in fact is (made of different kinds of rock, for example, or housing different flora and fauna); and we can even imagine mountains, beside crystal fountains, where the hens lay soft-boiled eggs. We can have conflicting thoughts about one and the same object, even if we don’t realize it: a mountain climber might desire to climb K2 and have no desire to scale Chogori, even though Chogori is K2. Franz Brentano first called attention to this set of phenomena in his Psychology from an Empirical Standpoint (1874/1973). It has come to be called intentionality. In fact, Brentano described intentionality as the mark of the mental: a feature had by all and only mental states.

经过适当编程的计算机是否也具有意向性？从表面上看，计算机具有“代表”其他事物的内部状态似乎是显而易见和微不足道的。根据第一部分中的某些观点，计算机是根据其操纵符号的能力来定义的，而符号通常代表事物。但在这里，我们必须小心，不要混淆状态可以代表其他事物的两种重要不同含义。一方面，我们可以将路标、书中的文字或图表上的线条说成是“代表”其他事物，即有人可以看到它们并解读它们的含义。这被称为衍生意向性，正是因为它的存在需要观察者为原本无意识的实体（标志上的油漆、页面上的印刷品等）提供意义。路标、书籍和图表本身并不知道它们的含义。相比之下，原始意向性是持续挑战的根源：我们如何才能构建系统，让世界本身就具有意义，而不是由解释者提供意义？更一般地讲，物理世界如何才能包含以这种更强的意义来代表事物的系统？本部分的文章阐述了人工智能 (AI) 面临的这一挑战，并对可能的解决方案进行了限制。

Could an appropriately programmed computer also have intentionality? On the face of it, it might seem obvious and trivial that computers have internal states that “represent” other things. According to some of the views in part I, computers are defined in terms of their ability to manipulate symbols, and symbols usually represent things. But here we must be careful not to confuse two importantly different senses in which states can be about other things. On the one hand, we can talk about road signs, words in a book, or lines on a graph as “representing” other things, in the sense that someone could look at them and decipher their meanings. This has come to be called derivative intentionality, precisely because it requires for its existence an observer who supplies meaning to what would otherwise be mindless entities (paint on a sign, print on a page, etc.). The road signs, books, and graphs don’t themselves have a clue of what they are about. Original intentionality, in contrast, is the source of the persistent challenge: How can we build systems that imbue the world with meaning themselves rather than having it supplied by an interpreter? More generally, how can a physical world contain systems that represent things in this stronger sense? The essays in this part articulate this challenge for artificial intelligence (AI) and place constraints on possible solutions.

在第 11 章中，丹尼特认为，不应从系统的组成部分或它们如何相互作用的角度来理解意向性，而应从组织中产生的行为的复杂性的角度来理解。根据他的说法，意向系统是这样一种系统，我们通过采取“意向立场”获得相当大的预测优势；也就是说，通过将系统视为具有信念、欲望、情感（等等），并预测系统将根据这些信念和欲望或多或少地理性地行事。这种观点将意向系统与仅仅是生物或物理的系统区分开来（例如）有些模糊。对于某些人来说，丹尼特的观点似乎将系统的意向性留在“旁观者的眼睛”中，因此提出了一个问题：物理世界如何可能首先包含采取立场的生物。事实上，人们可能会认为，按照第 2 章中 Haugeland 的说法，意向性就是采取立场的问题。如果是这样，它就不能简化为采取立场。

In chapter 11, Dennett argues that intentionality should not be understood in terms of the components of a system or how they interact, but in terms of the complexity of behavior that arises from their organization. An intentional system, by his account, is a system for which we gain considerable predictive leverage by taking the “intentional stance”; that is, by treating the system as if it has beliefs, desires, emotions (etc.) and predicting that the system will behave more or less rationally in light of those beliefs and desires. This view leaves the line between intentional systems and those that are merely biological or physical (for example) somewhat blurry. For some, Dennett’s view seems to leave the intentionality of the system in the “eye of the beholder,” and so raises the question of how it is possible for the physical world to contain creatures that take stances in the first place. Indeed, one might think, following Haugeland in chapter 2, that intentionality is a matter of taking a stance. If so, it can’t be reduced to stance-taking.

在第 12 章中，Searle 指出，理性可描述的行为，或者任何单纯的符号操作本身，都不足以体现意向性。他的“中文房间”思想实验已成为意向性挑战的领域定义性表达。它据称表明，即使一个使用复杂符号操作通过图灵测试的系统实际上也可能不理解它在说什么。他考虑了许多应对这一挑战的尝试，但最终都拒绝了，他最终认为最初的意向性（或理解）源自尚未发现的大脑独特生化特性。

In chapter 12, Searle argues that rationally describable action, or, indeed, any mere symbol manipulation by itself, will not suffice for intentionality. His Chinese Room thought experiment has become a field-defining articulation of the challenge of intentionality. It putatively shows that even a system that passes the Turing test using sophisticated symbol manipulation might not, in fact, understand what it is talking about. He considers many attempts to meet that challenge and rejects them all, holding ultimately that original intentionality (or understanding) emerges from the as-yet-undiscovered, unique biochemical properties of the brain.

在第 13 章中，博登为后来被称为“系统回应”的 Searle 论证进行了辩护。根据这种回应风格，Searle 的论证调查了一个据称具有意向性的系统的各个部分，在其各个部分中没有发现意向性，并错误地得出结论，认为整个系统不了解正在发生的事情。从这个角度来看，Searle 犯了一种组合谬误：假设整体的属性必须由其各个组成部分单独拥有。正如下文所述，这一回应引起了相当大的争议。

In chapter 13, Boden defends what has come to be called the “systems reply” to Searle’s argument. According to this style of reply, Searle’s argument investigates the parts of a purportedly intentional system, finds no intentionality in its individual parts, and wrongly concludes that the system as a whole does not understand what is going on. Searle, in this view, commits a kind of compositional fallacy: assuming that the property of the whole must be possessed by its components taken individually. This reply has been the subject of considerable controversy, as discussed next.

与这种反对意见一致，弗朗西斯·伊根在第 14 章中指出，计算系统是可定义的，可以严格根据其实例化的数学函数进行个体化。尽管如此，当我们用计算组件来解释系统解决问题的能力时，解释分为两部分：一部分是用计算系统的数学函数来狭义地描述计算系统；另一部分以明确计算机状态如何跟踪世界特征的方式“阐释”计算描述，以便在执行数学函数时解决问题。当然，塞尔认为也存在这样的阐释，并否认这足以满足原始意向性。额外的数学结构是否足以应对挑战仍是一个悬而未决的问题。

Consistent with this kind of objection, Francis Egan argues in chapter 14 that computational systems are definable and can be individuated strictly on the basis of the mathematical functions that they instantiate. Nonetheless, when we explain a capacity of a system to solve a problem in terms of its computational components, the explanation has two parts: one that describes the computational system narrowly in terms of the mathematical function that it computes, and one that “glosses” that computational description in ways that make explicit how the states of the computer track features of the world such that, in executing the mathematical function, it would solve the problem. Yet, of course, Searle thinks that there is such a gloss too and denies that this suffices for original intentionality. Whether the presence of additional mathematical structure is sufficient to address the challenge remains an open question.

11.1 有意策略及其运作方式

11.1  The intentional strategy and how it works

策略有很多，有好的，有坏的。例如，这里有一个预测某人未来行为的策略：确定此人的出生日期和时间，然后将这个简单的数据输入一个或另一个占星算法，以预测此人的前景。这种策略流行得令人遗憾。它之所以流行得令人遗憾，只是因为我们有充分的理由相信它不起作用（pace Feyerabend，1978）。当占星预测成真时，这纯粹是运气好，或者是由于预言非常模糊或含糊，几乎任何可能发生的情况都可以被解释为证实它。但假设占星策略确实对某些人很有效。我们可以把这些人称为占星系统——其行为事实上可以通过占星策略预测的系统。如果有这样的人，这样的占星系统，我们就会比大多数人更关心占星策略是如何运作的——也就是说，我们会对占星术的规则、原则或方法感兴趣。我们可以通过询问占星家、阅读他们的书籍和观察他们的实际行动来了解策略是如何运作的。但我们也会好奇它为什么有效。我们可能会发现占星家对后一个问题没有任何有用的意见——他们要么没有关于它为什么有效的理论，要么他们的理论纯粹是胡说八道。有一个好的策略是一回事，知道它为什么有效又是另一回事。

There are many strategies, some good, some bad. Here is a strategy, for instance, for predicting the future behavior of a person: determine the date and hour of the person’s birth and then feed this modest datum into one or another astrological algorithm for generating predictions of the person’s prospects. This strategy is deplorably popular. Its popularity is deplorable only because we have such good reasons for believing that it does not work (pace Feyerabend, 1978). When astrological predictions come true this is sheer luck, or the result of such vagueness or ambiguity in the prophecy that almost any eventuality can be construed to confirm it. But suppose the astrological strategy did in fact work well on some people. We could call those people astrological systems—systems whose behavior was, as a matter of fact, predictable by the astrological strategy. If there were such people, such astrological systems, we would be more interested than most of us in fact are in how the astrological strategy works—that is, we would be interested in the rules, principles, or methods of astrology. We could find out how the strategy works by asking astrologers, reading their books, and observing them in action. But we would also be curious about why it worked. We might find that astrologers had no useful opinions about this latter question—they either had no theory of why it worked or their theories were pure hokum. Having a good strategy is one thing; knowing why it works is another.

然而，据我们所知，占星术系统这一类是空洞的；因此占星术策略仅作为一种社会好奇心而引起人们的兴趣。其他策略则具有更好的信誉。考虑物理策略或物理立场；如果你想预测一个系统的行为，确定它的物理构成（可能一直到微观物理层面）和对它的冲击的物理性质，并利用你对物理定律的了解来预测任何输入的结果。这是拉普拉斯预测宇宙中一切事物的整个未来的宏伟而不切实际的策略；但它有更温和、更局部、更实际可用的版本。实验室里的化学家或物理学家可以使用这种策略来预测奇异材料的行为，但同样，厨房里的厨师可以预测锅在炉子上放得太久的影响。这种策略并不总是实用的，但它在原则上总是有效的，这是物理科学的教条。 （我忽略了量子物理亚原子不确定性引起的轻微复杂情况。）

So far as we know, however, the class of astrological systems is empty; so the astrological strategy is of interest only as a social curiosity. Other strategies have better credentials. Consider the physical strategy, or physical stance; if you want to predict the behavior of a system, determine its physical constitution (perhaps all the way down to the microphysical level) and the physical nature of the impingements upon it, and use your knowledge of the laws of physics to predict the outcome for any input. This is the grand and impractical strategy of Laplace for predicting the entire future of everything in the universe; but it has more modest, local, actually usable versions. The chemist or physicist in the laboratory can use this strategy to predict the behavior of exotic materials, but equally the cook in the kitchen can predict the effect of leaving the pot on the burner too long. The strategy is not always practically available, but that it will always work in principle is a dogma of the physical sciences. (I ignore the minor complications raised by the subatomic indeterminacies of quantum physics.)

无论如何，有时更有效的做法是从物理立场转向我所说的设计立场，在设计立场中，人们忽略物体物理构造的实际（可能很混乱）细节，并假设它具有某种设计，预测它在各种情况下的行为将按照设计的方式运行。例如，大多数计算机用户根本不知道哪些物理原理决定了计算机的高度可靠性，因此是可预测的。但是，如果他们很好地了解计算机的设计目的（在许多可能的抽象层次中的任何一个层次上对其操作的描述），他们就可以非常准确和可靠地预测其行为，只有在发生物理故障时才会出现偏差。不那么引人注目的是，几乎任何人都可以根据对闹钟外观的最随意的检查来预测闹钟何时响起。人们不知道或不想知道闹钟是弹簧上弦、电池驱动、太阳能驱动、由黄铜轮和宝石轴承还是硅片制成——人们只是假设闹钟的设计使得闹钟会在设定的时间响起，闹钟会在设定的时间响起，闹钟会一直运行到设定的时间甚至更久，闹钟的设计运行或多或少是准确的，等等。为了更准确、更详细地预测闹钟的设计立场，人们必须将其设计描述到不太抽象的层次；例如，描述齿轮的层次，但未指定其材料。

Sometimes, in any event, it is more effective to switch from the physical stance to what I call the design stance, where one ignores the actual (possibly messy) details of the physical constitution of an object, and, on the assumption that it has a certain design, predicts that it will behave as it is designed to behave under various circumstances. For instance, most users of computers have not the foggiest idea what physical principles are responsible for the computer’s highly reliable, and hence predictable, behavior. But if they have a good idea of what the computer is designed to do (a description of its operation at any one of the many possible levels of abstraction), they can predict its behavior with great accuracy and reliability, subject to disconfirmation only in the cases of physical malfunction. Less dramatically, almost anyone can predict when an alarm clock will sound on the basis of the most casual inspection of its exterior. One does not know or care to know whether it is spring wound, battery driven, sunlight powered, made of brass wheels and jewel bearings or silicon chips—one just assumes that it is designed so that the alarm will sound when it is set to sound, and it is set to sound where it appears to be set to sound, and the clock will keep on running until that time and beyond, and is designed to run more or less accurately, and so forth. For more accurate and detailed design stance predictions of the alarm clock, one must descend to a less abstract level of description of its design; for instance, to the level at which gears are described, but their material is not specified.

当然，只有系统的设计行为才是可以从设计角度预测的。如果你想预测闹钟在充满液氦时的行为，请回到物理角度。不仅是人工制品，而且许多生物物体（植物和动物、肾脏和心脏、雄蕊和雌蕊）的行为方式都可以从设计角度预测。它们不仅是物理系统，而且是设计系统。

Only the designed behavior of a system is predictable from the design stance, of course. If you want to predict the behavior of an alarm clock when it is pumped full of liquid helium, revert to the physical stance. Not just artifacts but also many biological objects (plants and animals, kidneys and hearts, stamens and pistils) behave in ways that can be predicted from the design stance. They are not just physical systems but designed systems.

有时甚至设计立场在实践中也难以实现，这时人们可以采取另一种立场或策略：意向立场。其工作原理如下：首先，你决定将要预测其行为的对象视为理性主体；然后，根据主体在世界中的地位和目的，弄清楚主体应该拥有哪些信念。然后，根据同样的考虑，弄清楚主体应该拥有哪些愿望，最后，你预测这个理性主体将根据其信念采取行动，进一步实现其目标。在许多情况下（但不是所有情况下），从所选的信念和愿望集合中进行一些实际推理，将得出主体应该做什么的决定；这就是你预测主体将要做的事情。

Sometimes even the design stance is practically inaccessible, and then there is yet another stance or strategy one can adopt: the intentional stance. Here is how it works: first you decide to treat the object whose behavior is to be predicted as a rational agent; then you figure out what beliefs that agent ought to have, given its place in the world and its purpose. Then you figure out what desires it ought to have, on the same considerations, and finally you predict that this rational agent will act to further its goals in the light of its beliefs. A little practical reasoning from the chosen set of beliefs and desires will in many—but not all—instances yield a decision about what the agent ought to do; that is what you predict the agent will do.

稍加阐述，这个策略就变得清晰起来。首先考虑一下我们如何向彼此灌输信念。以下是一些真理：受保护的人往往无知；如果你让某人接触某事，他就会知道一切。一般来说，似乎，我们开始相信我们所能了解的周围世界部分的所有真相。接触x——也就是说，在适当的一段时间内与x进行感官对抗——通常是了解（或拥有真实信念）x的充分条件。正如我们所说，我们开始了解周围的一切。这种接触通常只足以获得知识，但这并不是表面上看起来的一个巨大的逃生舱；我们面对接触时接受异常无知的门槛相当高。“我不知道枪里装了子弹”，一个在场、目睹并且清醒地装弹的人说，这遭到了各种各样的彻底怀疑，只有最离奇的支持故事才能压倒一切。

The strategy becomes clearer with a little elaboration. Consider first how we go about populating each other’s heads with beliefs. A few truisms: sheltered people tend to be ignorant; if you expose someone to something he comes to know all about it. In general, it seems, we come to believe all the truths about the parts of the world around us we are put in a position to learn about. Exposure to x—that is, sensory confrontation with x over some suitable period of time—is the normally sufficient condition for knowing (or having true beliefs) about x. As we say, we come to know all about the things around us. Such exposure is only normally sufficient for knowledge, but this is not the large escape hatch it might appear; our threshold for accepting abnormal ignorance in the face of exposure is quite high. “I didn’t know the gun was loaded”, said by one who was observed to be present, sighted, and awake during the loading, meets with a variety of utter skepticism that only the most outlandish supporting tale could overwhelm.

当然，我们不会学习或记住感官历史带给我们的所有事实。尽管有“知道一切”这个短语，但我们通常知道的只是感官历史带给我们的所有相关事实。我通常不知道我所居住的房间里戴眼镜的人和穿裤子的人的比例，但如果我对此感兴趣，那就很容易学会。这不仅仅是因为我周围的一些事实低于我的辨别阈值或超出了我的记忆的整合和保持能力（例如在场所有人的身高，以英寸为单位），而且许多完全可检测、可理解、可记忆的事实对我来说毫无兴趣，因此我不会相信它们。因此，在意向策略中归因信念的一条规则是：将系统迄今为止的经验所提供的与系统利益（或愿望）相关的所有事实归因于信念。这条规则会导致过度归因——因为我们都有点健忘，即使是重要的事情。它也未能捕捉到众所周知的错误信念。但是，错误信念（任何错误信念）的归因需要特殊的谱系，而这种谱系主要存在于真实信念中。两个典型案例：S（错误地）相信p，因为S（真正地）相信琼斯告诉他p，琼斯很聪明，琼斯无意欺骗他，……等等。第二种情况：S（错误地）相信酒吧凳上有一条蛇，因为S（真正地）相信他似乎看到了酒吧凳上有一条蛇，他自己坐在离他看到的酒吧凳不到一码远的酒吧里，等等。虚假必须从某个地方开始：种子可能种在幻觉、错觉、正常的简单错误感知、记忆衰退或故意欺诈中；但所收获的错误信念却是在真实信念的培养基中生长的。

Of course we do not come to learn or remember all the truths our sensory histories avail us. In spite of the phrase “know all about”, what we come to know, normally, are only all the relevant truths our sensory histories avail us. I do not typically come to know the ratio of spectacle-wearing people to trousered people in a room I inhabit, though if this interested me, it would be readily learnable. It is not just that some facts about my environment are below my thresholds of discrimination or beyond the integration and holding power of my memory (such as the height in inches of all the people present), but that many perfectly detectable, graspable, memorable facts are of no interest to me and hence do not come to be believed by me. So one rule for attributing beliefs in the intentional strategy is this: attribute as beliefs all the truths relevant to the system’s interests (or desires) that the system’s experience to date has made available. This rule leads to attributing somewhat too much—since we all are somewhat forgetful, even of important things. It also fails to capture the false beliefs we are all known to have. But the attribution of false belief, any false belief, requires a special genealogy, which will be seen to consist in the main in true beliefs. Two paradigm cases: S believes (falsely) that p, because S believes (truly) that Jones told him that p, that Jones is pretty clever, that Jones did not intend to deceive him,…and so on. Second case: S believes (falsely) that there is a snake on the barstool, because S believes (truly) that he seems to see a snake on the barstool, is himself sitting in a bar not a yard from the barstool he sees, and so forth. The falsehood has to start somewhere: the seed may be sown in hallucination, illusion, a normal variety of simple misperception, memory deterioration, or deliberate fraud, for instance; but the false beliefs that are reaped grow in a culture medium of true beliefs.

然后，在讨论信仰归因时，人们经常关注那些神秘而复杂的信仰，无论是真还是假。天知道，它们并不是直接来自接触平凡的事物和事件，但它们的归因需要从已经归因的大量信仰中追溯出一条主要为好的论据或推理的谱系。因此，有意策略的一个含义是，真正的信徒主要相信真理。如果有人能设计出一种公认的方法来对信仰进行个体化和计数（我非常怀疑），我们就会发现，除了最小的部分（比如说，不到百分之十）之外，一个人的所有信仰都可以在我们的第一条规则下归因^。1

Then there are the arcane and sophisticated beliefs, true and false, that are so often at the focus of attention in discussions of belief attribution. They do not arise directly, goodness knows, from exposure to mundane things and events, but their attribution requires tracing out a lineage of mainly good argument or reasoning from the bulk of beliefs already attributed. An implication of the intentional strategy, then, is that true believers mainly believe truths. If anyone could devise an agreed-upon method of individuating and counting beliefs (which I doubt very much), we would see that all but the smallest portion (say, less than ten percent) of a person’s beliefs were attributable under our first rule.¹

请注意，此规则是派生规则，是基本规则的详细说明和进一步规范：归因于系统应当具有的那些信念。还请注意，此规则与欲望归因相互作用。我们如何归因于欲望（偏好、目标、兴趣），并以此为基础形成信念列表？我们归因于系统应当具有的欲望。这是基本规则。它规定，首先，我们要将人们所熟悉的最高或最基本的欲望列表归因于人们：生存、无痛、食物、舒适、生育、娱乐。引用其中任何一个欲望通常都会终止“为什么？”的推理游戏。人们不需要别有用心来渴望舒适或快乐或延长自己的存在。欲望归因的派生规则与信念归因相互作用。显然，我们有一条规则：将欲望归因于系统认为对它有益的那些事物。更具体一点，将系统认为是最佳手段的欲望归因于它所希望的其他目的。因此，像归因错误信念一样，归因怪异和有害的欲望需要特殊的故事。

Note that this rule is a derived rule, an elaboration and further specification of the fundamental rule: attribute those beliefs the system ought to have. Note also that the rule interacts with the attribution of desires. How do we attribute the desires (preferences, goals, interests) on whose basis we will shape the list of beliefs? We attribute the desires the system ought to have. That is the fundamental rule. It dictates, on a first pass, that we attribute the familiar list of highest, or most basic, desires to people: survival, absence of pain, food, comfort, procreation, entertainment. Citing any one of these desires typically terminates the “Why?” game of reason giving. One is not supposed to need an ulterior motive for desiring comfort or pleasure or the prolongation of one’s existence. Derived rules of desire attribution interact with belief attributions. Trivially, we have the rule: attribute desires for those things a system believes to be good for it. Somewhat more informatively, attribute desires for those things a system believes to be best means to other ends it desires. The attribution of bizarre and detrimental desires thus requires, like the attribution of false beliefs, special stories.

当我们考虑根据言语行为将欲望归因于什么时，信念和欲望之间的相互作用变得更加棘手。用语言表达欲望的能力打开了欲望归因的闸门。“我想要一个两个鸡蛋的蘑菇煎蛋卷、一些法式面包和黄油，以及半瓶略微冰镇的勃艮第白葡萄酒。”在没有这种口头声明的情况下，人们怎么能开始将欲望归因于如此具体的东西呢？事实上，一个生物如何在没有语言帮助的情况下产生如此具体的欲望？语言使我们能够表达非常具体的欲望，但它也迫使我们偶尔承诺满足比我们本来有理由努力满足的任何事物更严格的满足条件的欲望。因为为了得到你想要的东西，你经常必须说出你想要的东西，而因为你经常不能在不说出比你先前的意思更具体的东西的情况下说出你想要的东西，所以你经常最终向别人提供证据（最好的证据，你不经意间说的话）来表明你想要的东西或事态比你能满足的要具体得多——或者比已经满足你的要好得多，因为一旦你声明了，作为一个言而有信的人，你就会对满足你所声明的愿望而不是其他愿望感兴趣。

The interaction between belief and desire becomes trickier when we consider what desires we attribute on the basis of verbal behavior. The capacity to express desires in language opens the floodgates of desire attribution. “I want a two-egg mushroom omelet, some French bread and butter, and a half bottle of lightly chilled white Burgundy.” How could one begin to attribute a desire for anything so specific in the absence of such verbal declaration? How, indeed, could a creature come to contract such a specific desire without the aid of language? Language enables us to formulate highly specific desires, but it also forces us on occasion to commit ourselves to desires altogether more stringent in their conditions of satisfaction than anything we would otherwise have any reason to endeavor to satisfy. Since in order to get what you want you often have to say what you want, and since you often cannot say what you want without saying something more specific than you antecedently mean, you often end up giving others evidence (the very best of evidence, your unextorted word) that you desire things or states of affairs far more particular than would satisfy you—or better, than would have satisfied you, for once you have declared, being a man of your word, you acquire an interest in satisfying exactly the desire you declared and no other.

您可能会反对别人要求您明确说明自己的愿望，但事实上，在日常生活中，我们都被社会化地接受类似的要求——甚至到了没有注意到的地步，当然也不会感到被压迫的地步。我之所以详细讨论这个问题，是因为它与信仰领域有相似之处，我们的语言环境永远迫使我们给出或让步于精确的言语表达，而这些信念缺乏言语化赋予它们的硬边（见 Dennett 1969，第 184-85 页；Dennett 1978b）。如果只关注这种社会力量的结果，而忽略其扭曲效应，人们很容易被误导，认为信仰和愿望显然就像储存在头脑中的句子。作为使用语言的生物，我们不可避免地会经常相信某个特定的、实际表述、拼写和标点的句子是真实的，并且在其他情况下我们希望这样的句子成真；但这些都是信仰和愿望的特殊情况，因此可能不是整个领域的可靠模型。

You might well object to having such a specification of desire demanded of you, but in fact we are all socialized to accede to similar requirements in daily life—to the point of not noticing it, and certainly not feeling oppressed by it. I dwell on this because it has a parallel in the realm of belief, where our linguistic environment is forever forcing us to give—or concede—precise verbal expression to convictions that lack the hard edges verbalization endows them with (see Dennett 1969, pp.184–85; Dennett 1978b). By concentrating on the results of this social force, while ignoring its distorting effect, one can easily be misled into thinking that it is obvious that beliefs and desires are rather like sentences stored in the head. Being language-using creatures, it is inevitable that we should often come to believe that some particular, actually formulated, spelled, and punctuated sentence is true, and that on other occasions we should come to want such a sentence to come true; but these are special cases of belief and desire and as such may not be reliable models for the whole domain.

至此，关于意向性策略中信念和欲望归因的原则就说得够多了。那么人们赋予意向系统的合理性又如何呢？人们从完美合理性的理想开始，并根据情况向下修正。也就是说，人们从这样的假设开始：人们相信他们信念的所有含义，并且不相信任何相互矛盾的信念对。这不会造成实际的混乱问题（例如无限多的含义），因为人们只关心确保所预测的系统足够合理，能够得出与其当前行为困境相关的特定含义。非理性的情况，或有限强大的推理能力，会引发特别棘手的解释问题，在此我将暂且不谈这些问题（参见 Dennett，1981；Cherniak，1986）。

That is enough, on this occasion, about the principles of belief and desire attribution to be found in the intentional strategy. What about the rationality one attributes to an intentional system? One starts with the ideal of perfect rationality and revises downward as circumstances dictate. That is, one starts with the assumption that people believe all the implications of their beliefs and believe no contradictory pairs of beliefs. This does not create a practical problem of clutter (infinitely many implications, for instance), for one is interested only in ensuring that the system one is predicting is rational enough to get to the particular implications that are relevant to its behavioral predicament of the moment. Instances of irrationality, or of finitely powerful capacities of inferences, raise particularly knotty problems of interpretation, which I will set aside on this occasion (see Dennett, 1981; Cherniak, 1986).

因为我想从对策略的描述转向对其使用的问题。人们真的会使用这种策略吗？是的，一直都在使用。也许有一天会存在其他归因信念和愿望以及预测行为的策略，但这是目前我们唯一知道的策略。它什么时候起作用？它几乎一直对人们有效。为什么允许牛津大学的各个学院在他们认为合适的时间创建和授予学位不是一个好主意？答案是一个很长的故事，但很容易得出。而且人们会就主要观点达成广泛共识。我们很容易想到人们会采取这样的行为，以给别人提供这样做的理由，从而给别人提供这样做的理由……创造我们不希望出现的情况。我们对有意策略的使用是如此习惯和毫不费力，以至于它在塑造我们对人的期望方面所起的作用很容易被忽视。这种策略在大多数情况下也适用于大多数其他哺乳动物。例如，你可以用它来设计更好的陷阱来捕捉哺乳动物，通过推理动物对各种事物的了解或信念、它的喜好和想要避免的东西。这一策略适用于鸟类、鱼类、爬行动物、昆虫和蜘蛛，甚至适用于像蛤蜊这样低等而缺乏进取心的生物（一旦蛤蜊认为有危险，它就不会放松紧握的壳，直到确信危险已经过去）。它也适用于一些人工制品：下棋的计算机不会吃掉你的骑士，因为它知道接下来的走法会导致失去车，它不希望这种情况发生。更温和地说，恒温器会在相信房间已达到所需温度时立即关闭锅炉。

For I want to turn from the description of the strategy to the question of its use. Do people actually use this strategy? Yes, all the time. There may someday be other strategies for attributing belief and desire and for predicting behavior, but this is the only one we all know now. And when does it work? It works with people almost all the time. Why would it not be a good idea to allow individual Oxford colleges to create and grant academic degrees whenever they saw fit? The answer is a long story, but very easy to generate. And there would be widespread agreement about the major points. We have no difficulty thinking of the reasons people would then have for acting in such ways as to give others reasons for acting in such ways as to give others reasons for…creating a circumstance we would not want. Our use of the intentional strategy is so habitual and effortless that the role it plays in shaping our expectations about people is easily overlooked. The strategy also works on most other mammals most of the time. For instance, you can use it to design better traps to catch those mammals, by reasoning about what the creature knows or believes about various things, what it prefers, what it wants to avoid. The strategy works on birds, and on fish, and on reptiles, and on insects and spiders, and even on such lowly and unenterprising creatures as clams (once a clam believes there is danger about, it will not relax its grip on its closed shell until it is convinced that the danger has passed). It also works on some artifacts: the chess-playing computer will not take your knight because it knows that there is a line of ensuing play that would lead to losing its rook, and it does not want that to happen. More modestly, the thermostat will turn off the boiler as soon as it comes to believe the room has reached the desired temperature.

这一策略甚至适用于植物。在春末风暴频发的地区，你应该种植那些对春天到来特别谨慎的苹果品种——当然，春天正是它们开花的时候。这一策略甚至适用于闪电等无生命且明显非人为的现象。一位电工曾向我解释他是如何保护我的地下水泵免受雷击损害的：他说，闪电总是想找到最佳接地方式，但有时它会被诱骗走次优路径。你可以通过让另一条更好的路径更容易被闪电发现来保护水泵。

The strategy even works for plants. In a locale with late spring storms, you should plant apple varieties that are particularly cautious about concluding that it is spring—which is when they want to blossom, of course. It even works for such inanimate and apparently undesigned phenomena as lightning. An electrician once explained to me how he worked out how to protect my underground water pump from lightning damage: lightning, he said, always wants to find the best way to ground, but sometimes it gets tricked into taking second-best paths. You can protect the pump by making another, better path more obvious to the lightning.

11.2 真正的信徒作为意向系统

11.2  True believers as intentional systems

现在，很明显，这是“严肃的”信念归因、可疑的信念归因、教学上有用的隐喻、说话方式，甚至更糟的彻头彻尾的欺诈行为的杂乱无章的集合。下一个任务似乎是区分那些真正具有信念和欲望的意向系统和那些我们可能觉得方便对待的、好像它们具有信念和欲望的系统。但那将是一项西西弗斯式的劳动，否则就会被命令终止。更好地理解信念现象始于这样的观察：即使在最糟糕的情况下，即使我们最确信该策略出于错误的原因而起作用，它仍然确实起作用，至少有一点。这是一个有趣的事实，它将这类对象（意向系统类）与该策略永远不起作用的对象类区分开来。但事实如此吗？我们对意向系统的定义是否排除了任何对象？例如，似乎可以将这个教室里的讲台解释为一个有意识的系统，完全理性，相信它目前位于文明世界的中心（你们中的一些人可能也这么认为），并且最希望的是留在那个中心。这样一个拥有信仰和欲望的理性主体应该做什么？显然，留在原地——这正是讲台所做的。我从有意识的立场准确地预测了讲台的行为，那么它是一个有意识的系统吗？如果是，那么任何事情都是有意识的。

Now clearly this is a motley assortment of “serious” belief attributions, dubious belief attributions, pedagogically useful metaphors, façons de parler, and, perhaps worse, outright frauds. The next task would seem to be distinguishing those intentional systems that really have beliefs and desires from those we may find it handy to treat as if they had beliefs and desires. But that would be a Sisyphean labor, or else would be terminated by fiat. A better understanding of the phenomenon of belief begins with the observation that even in the worst of these cases, even when we are surest that the strategy works for the wrong reasons, it is nevertheless true that it does work, at least a little bit. This is an interesting fact, which distinguishes this class of objects, the class of intentional systems, from the class of objects for which the strategy never works. But is this so? Does our definition of an intentional system exclude any objects at all? For instance, it seems the lectern in this lecture room can be construed as an intentional system, fully rational, believing that it is currently located at the center of the civilized world (as some of you may also think), and desiring above all else to remain at that center. What should such a rational agent so equipped with belief and desire do? Stay put, clearly—which is just what the lectern does. I predict the lectern’s behavior, accurately, from the intentional stance, so is it an intentional system? If it is, anything at all is.

什么应该取消讲台的资格？首先，这种策略在这种情况下不适用，因为我们没有从中获得我们之前没有的预测能力。我们已经知道讲台要做什么——也就是什么也不做——并且以一种非常不合原则的方式调整信仰和愿望以适应它。然而，在人、动物或计算机的情况下，情况就不同了。在这些情况下，通常唯一可行的策略是有意策略；它赋予我们其他方法无法获得的预测能力。但是，有人会强调，这并不是本质上的差异，而只是反映了我们作为科学家的有限能力的差异。拉普拉斯全知物理学家可以预测计算机的行为——或者假设活体人体最终受物理定律支配——而无需采用设计或有意策略的冒险捷径方法。对于机械能力有限的人来说，对简单恒温器的有意解释是一种方便且基本无害的拐杖，但我们当中的工程师无需借助这种拟人化就可以完全掌握其内部运作。也许最聪明的工程师几乎不可能对更复杂的系统（如分时计算机系统或遥控太空探测器）保持清晰的概念，而不会陷入意向立场（将这些设备视为询问和告知、尝试和避免、想要和相信），但这只是人类认知弱点的更高级例子。我们不想基于如此多变和狭隘的理由将这些人工制品与真正的信徒（我们自己）归为一类，不是吗？从一位观察者的角度来看，某些人工制品、生物或人是信徒，但从另一位更聪明的观察者的角度来看，根本不是信徒，这难道不是无法容忍的吗？这将是解释主义的一个特别激进的版本，有些人认为我支持这种观点，因为我敦促从意向策略的成功的角度来看待这种信念。我必须承认，我提出的观点有时会引发这种解读，但现在我想阻止这种解读。是否采取意向立场的决定是自由的，但是，如果采取了该立场，其成功或失败的事实是完全客观的。

What should disqualify the lectern? For one thing, the strategy does not recommend itself in this case, for we get no predictive power from it that we did not antecedently have. We already knew what the lectern was going to do—namely nothing—and tailored the beliefs and desires to fit in a quite unprincipled way. In the case of people or animals or computers, however, the situation is different. In these cases often the only strategy that is at all practical is the intentional strategy; it gives us predictive power we can get by no other method. But, it will be urged, this is no difference in nature, but merely a difference that reflects upon our limited capacities as scientists. The Laplacean omniscient physicist could predict the behavior of a computer—or of a live human body, assuming it to be ultimately governed by the laws of physics—without any need for the risky, short-cut methods of either the design or intentional strategies. For people of limited mechanical aptitude, the intentional interpretation of a simple thermostat is a handy and largely innocuous crutch, but the engineers among us can quite fully grasp its internal operation without the aid of this anthropomorphizing. It may be true that the cleverest engineers find it practically impossible to maintain a clear conception of more complex systems, such as a time-sharing computer system or remote-controlled space probe, without lapsing into an intentional stance (and viewing these devices as asking and telling, trying and avoiding, wanting and believing), but this is just a more advanced case of human epistemic frailty. We would not want to classify these artifacts with the true believers—ourselves—on such variable and parochial grounds, would we? Would it not be intolerable to hold that some artifact or creature or person was a believer from the point of view of one observer, but not a believer at all from the point of view of another, cleverer observer? That would be a particularly radical version of interpretationism, and some have thought I espoused it in urging that belief be viewed in terms of the success of the intentional strategy. I must confess that my presentation of the view has sometimes invited that reading, but I now want to discourage it. The decision to adopt the intentional stance is free, but the facts about the success or failure of the stance, were one to adopt it, are perfectly objective.

一旦实施了有意策略，它就是一种非常强大的预测工具——这一事实在很大程度上被我们通常关注的、它产生可疑或不可靠结果的情况所掩盖。例如，考虑预测国际象棋比赛中的走法。人们可以看到，国际象棋之所以有趣，是因为对手的走法不可预测，除非在那些“被迫”走法的情况下——显然只有一种最佳走法——通常是所有可行走法中危害最小的一种。但是，当人们认识到在典型的国际象棋情况下，有很多完全合法且可用的走法，但只有少数——也许只有六种——有可取之处，因此根据有意策略，只有少数几步是高概率走法时，这种不可预测性就被放在了背景中。即使有意策略无法区分出概率最高的单一走法，它也会大大减少可行的选择数量。

Once the intentional strategy is in place, it is an extraordinarily powerful tool in prediction—a fact that is largely concealed by our typical concentration on the cases in which it yields dubious or unreliable results. Consider, for instance, predicting moves in a chess game. What makes chess an interesting game, one can see, is the unpredictability of one’s opponent’s moves, except in those cases where moves are “forced”—where there is clearly one best move—typically the least of the available evils. But this unpredictability is put in context when one recognizes that in the typical chess situation there are very many perfectly legal and hence available moves, but only a few—perhaps half a dozen—with anything to be said for them, and hence only a few high-probability moves according to the intentional strategy. Even when the intentional strategy fails to distinguish a single move with a highest probability, it can dramatically reduce the number of live options.

意图策略的相同特征在应用于“现实世界”案例时显而易见。众所周知，它无法预测股票交易员的确切买入和卖出决定，或者政客在预定演讲中会说出的确切词序。但是，对于稍微不那么具体的预测，人们的信心确实非常高：例如，特定交易员今天不会购买公用事业，或者政客会站在工会一边反对他的政党。从另一个角度来看，这种无法预测行为细粒度描述的能力是意图策略的优势之源，因为正是这种对实施细节的中立性，才允许人们在复杂情况下利用意图策略，例如，在连锁预测中（见 Dennett，1978a）。假设美国国务卿宣布他是克格勃的付费特工。这是多么史无前例的事件！其后果又是多么难以预测！但事实上，我们可以预测几十个虽然不是很有趣但却非常突出的后果，以及后果的后果。总统将与内阁其他成员商议，内阁将支持他决定在各种调查（精神病学和政治）结果出来之前解除国务卿的职务，所有这些都将在新闻发布会上报告给那些撰写报道的人，这些报道将在社论中被评论，而社论又会被那些给编辑写信的人阅读，等等。这些都不是大胆的预言，但请注意，它描述了时空中的因果关系弧，任何可以想象的物理学或生物学的实际扩展都无法预测这种因果关系弧。

The same feature of the intentional strategy is apparent when it is applied to “real world” cases. It is notoriously unable to predict the exact purchase and sell decisions of stock traders, for instance, or the exact sequence of words a politician will utter when making a scheduled speech. But one’s confidence can be very high indeed about slightly less specific predictions: that the particular trader will not buy utilities today, or that the politician will side with the unions against his party, for example. This inability to predict fine-grained descriptions of actions, looked at another way, is a source of strength for the intentional strategy, for it is this neutrality with regard to details of implementation that permits one to exploit the intentional strategy in complex cases, for instance, in chaining predictions (see Dennett, 1978a). Suppose the US Secretary of State were to announce he was a paid agent of the KGB. What an unparalleled event! How unpredictable its consequences! Yet in fact we can predict dozens of not terribly interesting but perfectly salient consequences, and consequences of consequences. The President would confer with the rest of the Cabinet, which would support his decision to relieve the Secretary of State of his duties pending the results of various investigations, psychiatric and political, and all this would be reported at a news conference to people who would write stories that would be commented upon in editorials that would be read by people who would write letters to the editors, and so forth. None of that is daring prognostication, but note that it describes an arc of causation in space-time that could not be predicted under any description by any imaginable practical extension of physics or biology.

几年前，罗伯特·诺齐克首次提出了一个反对意见，这让我们更清楚地看到了意向性策略的力量。他提出，假设一些拥有超凡智慧的生物（比如来自火星的生物）降临到我们身边，假设我们之于他们就像简单的恒温器之于聪明的工程师。也就是说，假设他们不需要意向性立场，甚至不需要设计立场，就能预测我们的行为的所有细节。我们可以假设他们是拉普拉斯超级物理学家，能够在微观物理层面理解华尔街的活动。我们看到的是经纪人和建筑物以及卖单和买单，而他们看到的是大量亚原子粒子在四处乱窜——他们是如此优秀的物理学家，以至于他们可以提前几天预测每天在标有“道琼斯工业平均指数收盘价”的纸带上会出现什么墨迹。他们可以预测他们观察到的所有各种运动物体的个体行为，而无需将它们中的任何一个视为有意识的系统。那么，我们可以说，从他们的角度来看，我们实际上根本不是信徒（就像一个简单的恒温器一样）吗？如果是这样，那么我们作为信徒的身份就不是客观的，而是旁观者眼中的事物——前提是旁观者与我们一样智力有限。

The power of the intentional strategy can be seen even more sharply with the aid of an objection first raised by Robert Nozick some years ago. Suppose, he suggested, some beings of vastly superior intelligence—from Mars, let us say—were to descend upon us, and suppose that we were to them as simple thermostats are to clever engineers. Suppose, that is, that they did not need the intentional stance—or even the design stance—to predict our behavior in all its detail. They can be supposed to be Laplacean super-physicists, capable of comprehending the activity on Wall Street, for instance, at the microphysical level. Where we see brokers and buildings and sell orders and bids, they see vast congeries of subatomic particles milling about—and they are such good physicists that they can predict days in advance what ink marks will appear each day on the paper tape labeled “Closing Dow Jones Industrial Average”. They can predict the individual behaviors of all the various moving bodies they observe without ever treating any of them as intentional systems. Would we be right then to say that from their point of view we really were not believers at all (any more than a simple thermostat is)? If so, then our status as believers is nothing objective, but rather something in the eye of the beholder—provided the beholder shares our intellectual limitations.

我们想象中的火星人也许能够通过拉普拉斯方法预测人类的未来，但如果他们不把我们看作有意识的系统，他们就会错过一些完全客观的东西：人类行为的模式，这些模式可以从有意识的立场描述，并且只能从这个立场描述，并且支持概括和预测。举一个具体的例子，火星人观察一位股票经纪人决定订购 500 股通用汽车股票。他们预测了他拨打电话时手指的确切动作，以及他发出订单时声带的确切振动。但是，如果火星人没有看到，无数不同的手指运动和声带振动模式——甚至无数不同个体的运动——可以取代实际的细节，而不会扰乱随后的市场运作，那么他们就没有看到他们所观察的世界中的真实模式。就像火花塞有无数种制作方法一样（除非人们意识到可以将各种不同的设备拧入这些插座而不会影响发动机的性能，否则人们就无法理解内燃机是什么），同样，订购 500 股通用汽车股票的方法也有很多，而且在社会插座中，其中一种方式产生的效果与其他方式几乎相同。社会枢纽也存在，人们选择哪条路取决于他们是否相信 p或渴望 A，而不取决于它们可能相似或不同的其他无限多种方式。

Our imagined Martians might be able to predict the future of the human race by Laplacean methods, but if they did not also see us as intentional systems, they would be missing something perfectly objective: the patterns in human behavior that are describable from the intentional stance, and only from that stance, and that support generalizations and predictions. Take a particular instance in which the Martians observe a stockbroker deciding to place an order for 500 shares of General Motors. They predict the exact motions of his fingers as he dials the phone and the exact vibrations of his vocal cords as he intones his order. But if the Martians do not see that indefinitely many different patterns of finger motions and vocal cord vibrations—even the motions of indefinitely many different individuals—could have been substituted for the actual particulars without perturbing the subsequent operation of the market, then they have failed to see a real pattern in the world they are observing. Just as there are indefinitely many ways of being a spark plug—and one has not understood what an internal combustion engine is unless one realizes that a variety of different devices can be screwed into these sockets without affecting the performance of the engine—so there are indefinitely many ways of ordering 500 shares of General Motors, and there are societal sockets in which one of these ways will produce just about the same effect as any other. There are also societal pivot points, as it were, where which way people go depends on whether they believe that p, or desire A, and does not depend on any of the other infinitely many ways they may be alike or different.

假设，让我们再进一步探讨一下火星人的幻想，一个火星人和一个地球人进行一场预测竞赛。地球人和火星人观察（并观察彼此观察）当地物理交换的某个部分。从地球人的角度来看，观察到的情况如下。加德纳太太的厨房里响起了电话铃声。她接了电话，说了这样的话：“哦，你好，亲爱的。你早点回家？一个小时内？带老板来吃饭？那么回家的路上买瓶酒，小心开车。”根据这一观察，我们的地球人预测，一辆大型金属橡胶轮胎车将在一小时内停在车道上，车里会有两个人，其中一个拿着一个纸袋，里面装着一瓶酒精。这个预测可能有点冒险，但从各方面来看都是一个不错的选择。火星人也做出了同样的预测，但他必须利用更多关于大量相互作用的信息，而据他所知，地球人对此一无所知。例如，在距离房屋五英里的A路口，车辆减速，否则就会与另一辆车相撞——火星人已经费尽心思计算了几百米的碰撞路线。地球人的表现看起来就像魔术一样！地球人怎么知道下车去商店拿瓶子的人会回来？在火星人绘制的路径经历了所有的变幻莫测、交叉和分支之后，地球人的预言得以实现，在任何人看来，如果没有故意的策略，这似乎就像萨迈拉之约的宿命论必然性一样奇妙和难以解释。宿命论者（例如占星家）相信，人类事务中存在着一种不可抗拒的模式，无论发生什么，这种模式都会自行发挥作用，也就是说，无论受害者如何策划和猜测，无论他们如何在锁链中扭来扭去。这些宿命论者错了，但他们几乎是对的。人类事务中存在着一种模式，这种模式并非完全不可抗拒，但却非常有力，吸收了可能被认为是随机的物理扰动和变化；这些是我们用理性主体的信念、欲望和意图来描述的模式。

Suppose, pursuing our Martian fantasy a little further, that one of the Martians were to engage in a predicting contest with an Earthling. The Earthling and the Martian observe (and observe each other observing) a particular bit of local physical transaction. From the Earthling’s point of view, this is what is observed. The telephone rings in Mrs. Gardner’s kitchen. She answers, and this is what she says: “Oh, hello dear. You’re coming home early? Within the hour? And bringing the boss to dinner? Pick up a bottle of wine on the way home then, and drive carefully.” On the basis of this observation, our Earthling predicts that a large metallic vehicle with rubber tires will come to a stop on the drive within one hour, disgorging two human beings, one of whom will be holding a paper bag containing a bottle containing an alcoholic fluid. The prediction is a bit risky, perhaps, but a good bet on all counts. The Martian makes the same prediction, but has to avail himself of much more information about an extraordinary number of interactions of which, so far as he can tell, the Earthling is entirely ignorant. For instance, the deceleration of the vehicle at intersection A, five miles from the house, without which there would have been a collision with another vehicle—whose collision course had been laboriously calculated over some hundreds of meters by the Martian. The Earthling’s performance would look like magic! How did the Earthling know that the human being who got out of the car and got the bottle in the shop would get back in? The coming true of the Earthling’s prediction, after all the vagaries, intersections, and branches in the paths charted by the Martian, would seem to anyone bereft of the intentional strategy as marvelous and inexplicable as the fatalistic inevitability of the appointment in Samarra. Fatalists—for instance, astrologers—believe that there is a pattern in human affairs that is inexorable, that will impose itself come what may, that is, no matter how the victims scheme and second-guess, no matter how they twist and turn in their chains. These fatalists are wrong, but they are almost right. There are patterns in human affairs that impose themselves, not quite inexorably but with great vigor, absorbing physical perturbations and variations that might as well be considered random; these are the patterns that we characterize in terms of the beliefs, desires, and intentions of rational agents.

毫无疑问，您已经注意到了我们思想实验中的一个严重缺陷，并因此而分心：火星人被认为将他的地球对手视为与他一样聪明的生物，可以与他交流，可以与他打赌，可以与他竞争。简而言之，火星人是一个有信念（比如他在预测中表达的信念）和欲望（比如赢得预测比赛的欲望）的生物。因此，如果火星人在一个地球人身上看到了这种模式，他怎么可能看不到其他地球人身上的模式呢？作为一点叙述，我们的例子可以通过假设我们的地球人巧妙地学习了火星语（通过 X 射线调制传播）并伪装成火星人来加强，依靠这些原本聪明的外星人的物种沙文主义，让他成为一个有意图的系统，而不会泄露他的同胞人类的秘密。这一补充可能会让我们摆脱故事中的糟糕转折，但可能会掩盖要得出的寓意：即，对自己和同类智能生物采取意向立场的不可避免性。这种不可避免性本身是与利益相关的；例如，完全有可能对包括自己在内的智能生物采取物理立场，但并不排除同时对自己和同类保持意向立场，例如，如果一个人打算学习他们所知道的东西（斯图尔特·汉普郡在许多著作中都有力地阐述了这一点）。我们或许可以假设我们的超级智能火星人未能将我们识别为意向系统，但我们不能假设他们缺乏必要的概念。2^如果他们观察、理论化、预测、交流，他们就会将自己视为意向系统。3有智能生物的地方，就必须有模式可供描述，无论我们是否愿意看到它们。

No doubt you will have noticed, and been distracted by, a serious flaw in our thought experiment: the Martian is presumed to treat his Earthling opponent as an intelligent being like himself, with whom communication is possible, a being with whom one can make a wager, against whom one can compete. In short, a being with beliefs (such as the belief he expressed in his prediction) and desires (such as the desire to win the prediction contest). So if the Martian sees the pattern in one Earthling, how can he fail to see it in the others? As a bit of narrative, our example could be strengthened by supposing that our Earthling cleverly learned Martian (which is transmitted by X-ray modulation) and disguised himself as a Martian, counting on the species-chauvinism of these otherwise brilliant aliens to permit him to pass as an intentional system while not giving away the secret of his fellow human beings. This addition might get us over a bad twist in the tale, but might obscure the moral to be drawn: namely, the unavoidability of the intentional stance with regard to oneself and one’s fellow intelligent beings. This unavoidability is itself interest relative; it is perfectly possible to adopt a physical stance, for instance, with regard to an intelligent being, oneself included, but not to the exclusion of maintaining at the same time an intentional stance with regard to oneself at a minimum, and one’s fellows if one intends, for instance, to learn what they know (a point that has been powerfully made by Stuart Hampshire in a number of writings). We can perhaps suppose our super-intelligent Martians fail to recognize us as intentional systems, but we cannot suppose them to lack the requisite concepts.² If they observe, theorize, predict, communicate, they view themselves as intentional systems.³ Where there are intelligent beings, the patterns must be there to be described, whether or not we care to see them.

认识到智能生物活动中可辨别的有意模式的客观现实很重要，但认识到模式中的不完整性和不完善性也很重要。客观事实是，有意策略虽然行之有效，但并不完美。没有人是完全理性、完全不会忘记、无所不知或不会疲劳、不会发生故障或设计缺陷的。这不可避免地会导致超出有意策略描述能力的情况，就像电话或汽车等人工制品受到物理损坏时，可能无法用该人工制品的正常设计术语来描述一样。如何绘制部分熔化的音频放大器的原理图接线图，或者如何描述故障计算机的程序状态？即使是最轻微和最常见的认知病理学情况——例如，人们似乎持有矛盾的信念，或者在自欺欺人——对意向策略的解释准则也无法对将哪些信念和愿望归因于一个人得出明确、稳定的结论。

It is important to recognize the objective reality of the intentional patterns discernible in the activities of intelligent creatures, but also important to recognize the incompleteness and imperfections in the patterns. The objective fact is that the intentional strategy works as well as it does, which is not perfectly. No one is perfectly rational, perfectly unforgetful, all-observant, or invulnerable to fatigue, malfunction, or design imperfection. This leads inevitably to circumstances beyond the power of the intentional strategy to describe, in much the same way that physical damage to an artifact, such as a telephone or an automobile, may render it indescribable by the normal design terminology for that artifact. How do you draw the schematic wiring diagram of an audio amplifier that has been partially melted, or how do you characterize the program state of a malfunctioning computer? In cases of even the mildest and most familiar cognitive pathology—where people seem to hold contradictory beliefs or to be deceiving themselves, for instance—the canons of interpretation of the intentional strategy fail to yield clear, stable verdicts about which beliefs and desires to attribute to a person.

现在，一个对信仰和欲望持强烈现实主义立场的人会声称，在这些情况下，所讨论的人确实有一些特定的信仰和欲望，而正如我所描述的，意向策略根本无法预测这些信仰和欲望。在我所提倡的较为温和的现实主义中，没有事实表明一个人在这些退化的情况下究竟有哪些信仰和欲望，但这并不是向相对主义或主观主义投降，因为何时以及为何没有事实本身就是一个客观事实问题。根据这种观点，人们甚至可以承认信仰归因的兴趣相对性，并承认鉴于不同文化的不同兴趣，例如，一种文化赋予某个成员的信仰和欲望可能与另一种文化赋予同一个人的信仰和欲望大不相同。但假设在特定情况下确实如此，那么还需要进一步的事实来说明每种相互竞争的意向策略在预测该人的行为方面效果如何。我们可以提前确定，对个人的任何有意解释都不会完美无缺，而且可能存在两种竞争方案，它们同样好，甚至比我们能想出的任何其他方案都好。这种情况本身就是事实。一种模式的客观存在（无论其缺陷如何）并不排除另一种模式的客观存在（无论其缺陷如何）。

Now a strong realist position on beliefs and desires would claim that in these cases the person in question really does have some particular beliefs and desires which the intentional strategy, as I have described it, is simply unable to divine. On the milder sort of realism I am advocating, there is no fact of the matter of exactly which beliefs and desires a person has in these degenerate cases, but this is not a surrender to relativism or subjectivism, for when and why there is no fact of the matter is itself a matter of objective fact. On this view one can even acknowledge the interest relativity of belief attributions and grant that given the different interests of different cultures, for instance, the beliefs and desires one culture would attribute to a member might be quite different from the beliefs and desires another culture would attribute to the very same person. But supposing that were so in a particular case, there would be the further facts about how well each of the rival intentional strategies worked for predicting the behavior of that person. We can be sure in advance that no intentional interpretation of an individual will work to perfection, and it may be that two rival schemes are about equally good, and better than any others we can devise. That this is the case is itself something about which there can be a fact of the matter. The objective presence of one pattern (with whatever imperfections) does not rule out the objective presence of another pattern (with whatever imperfections).

与意向性策略具有同等依据的截然不同的解释的隐患在理论上是重要的——或许可以说在形而上学上是重要的——但在实践中，一旦人们将注意力集中在我们所知的最大、最复杂的意向性系统：人类身上，这种隐患就可以忽略不计了^。4

The bogey of radically different interpretations with equal warrant from the intentional strategy is theoretically important—one might better say metaphysically important—but practically negligible once one restricts one’s attention to the largest and most complex intentional systems we know: human beings.⁴

到目前为止，我一直在强调我们与蛤蜊和恒温器的亲缘关系，以强调信念归因的逻辑地位，但现在是时候承认明显的差异并说明如何利用它们了。错误的说法仍然存在：成为一个真正的信徒只需要一个系统，其行为可以通过意向策略可靠地预测，因此真正相信p（对于任何命题 p）只需要一个意向系统，对于这个系统，p在最佳（最具预测性）解释中作为信念出现。但是，一旦我们把注意力转向真正有趣和多功能的意向系统，我们就会发现，这种看似肤浅和工具主义的信念标准对真正的信徒的内部构成施加了严格的限制，从而最终产生了一种强有力的信念。

Until now I have been stressing our kinship to clams and thermostats, in order to emphasize a view of the logical status of belief attribution, but the time has come to acknowledge the obvious differences and say what can be made of them. The perverse claim remains: all there is to being a true believer is being a system whose behavior is reliably predictable via the intentional strategy, and hence all there is to really and truly believing that p (for any proposition p) is being an intentional system for which p occurs as a belief in the best (most predictive) interpretation. But once we turn out attention to the truly interesting and versatile intentional systems, we see that this apparently shallow and instrumentalistic criterion of belief puts a severe constraint on the internal constitution of a genuine believer, and thus yields a robust version of belief after all.

以低级恒温器为例，它是可以想象到的能引起我们注意超过片刻的退化的意向系统案例。顺着这个笑话，我们或许同意赋予它大约六种不同信念和更少欲望的能力——它可以相信房间太冷或太热，锅炉开着或关着，如果它想让房间暖和点，它应该打开锅炉，等等。但这肯定是把太多的东西归咎于恒温器；例如，它没有热量或锅炉的概念。所以假设我们反解释它的信念和欲望：它可以相信 A 太 F 或 G，如果它想让 A 更 F，它应该做 K，等等。毕竟，通过将恒温控制机制连接到不同的输入和输出设备，它可以调节水箱中的水量，或者火车的速度，等等。它与热敏传感器和锅炉的连接过于贫乏，无法赋予其类似信仰的状态任何丰富的语义。

Consider the lowly thermostat, as degenerate a case of intentional system as could conceivably hold our attention for more than a moment. Going along with the gag, we might agree to grant it the capacity for about half a dozen different beliefs and fewer desires—it can believe the room is too cold or too hot, that the boiler is on or off, and that if it wants the room warmer it should turn on the boiler, and so forth. But surely this is imputing too much to the thermostat; it has no concept of heat or of a boiler, for instance. So suppose we de-interpret its beliefs and desires: it can believe the A is too F or G, and if it wants the A to be more F it should do K, and so forth. After all, by attaching the thermostatic control mechanism to different input and output devices, it could be made to regulate the amount of water in a tank, or the speed of a train, for instance. Its attachment to a heat sensitive transducer and a boiler is too impoverished a link to the world to grant any rich semantics to its belief-like states.

但是，假设我们随后丰富了这些依恋模式。假设我们给它提供多种了解温度的方法，例如。我们给它一只眼睛，可以区分房间里蜷缩着、瑟瑟发抖的人，还有一只耳朵，这样它就可以知道天气有多冷。我们给它一些地理知识，这样如果它知道它的时空位置是 12 月的温尼伯，它就可以得出结论，它可能在一个寒冷的地方。当然，要给它一个多用途、通用的视觉系统——而不仅仅是一个颤抖物体探测器——将需要其内部结构的巨大复杂性。假设我们还为系统提供了更多的行为多功能性：它选择锅炉燃料、从最便宜、最可靠的经销商处购买、检查密封条等等。这增加了内部复杂性的另一个维度；它为各个信念状态提供了更多的事情可做，实际上是通过提供更多不同的机会来从其他状态推导或推断它们，并通过提供更多不同的机会让它们作为进一步推理的前提。丰富设备与其所在世界之间的这些连接的累积效应是丰富其虚拟谓词 F 和 G 等的语义。我们添加的这些越多，我们的设备就越不适合用作除室温维持系统以外的任何系统的控制结构。更正式的说法是，随着我们增加这些复杂性，体现在形式系统内部状态中的难以区分的令人满意的模型类别变得越来越小；我们添加的越多，系统的语义就越丰富、要求就越高或越具体，直到最终我们达到在实践上（但从未在原则上）规定了唯一语义解释的系统（参见 Hayes，1979）。此时，我们说这个设备（或动物或人）对热量和这个房间等有信念，这不仅是因为系统在世界上的实际位置和操作，还因为我们无法想象它可以被放置在另一个可以工作的环境中（另见 Dennett，1982，1987a）。

But suppose we then enrich these modes of attachment. Suppose we give it more than one way of learning about the temperature, for instance. We give it an eye of sorts that can distinguish huddled, shivering occupants of the room and an ear so that it can be told how cold it is. We give it some facts about geography so that it can conclude that is probably in a cold place if it learns that its spatio-temporal location is Winnipeg in December. Of course giving it a visual system that is multipurpose and general—not a mere shivering-object detector—will require vast complications of its inner structure. Suppose we also give our system more behavioral versatility: it chooses the boiler fuel, purchases it from the cheapest and most reliable dealer, checks the weather stripping, and so forth. This adds another dimension of internal complexity; it gives individual belief-like states more to do, in effect, by providing more and different occasions for their derivation or deduction from other states, and by providing more and different occasions for them to serve as premises for further reasoning. The cumulative effect of enriching these connections between the device and the world in which it resides is to enrich the semantics of its dummy predicates, F and G and the rest. The more of this we add, the less amenable our device becomes to serving as the control structure of anything other than a room-temperature maintenance system. A more formal way of saying this is that the class of indistinguishably satisfactory models of the formal system embodied in its internal states gets smaller and smaller as we add such complexities; the more we add, the richer or more demanding or specific the semantics of the system, until eventually we reach systems for which a unique semantic interpretation is practically (but never in principle) dictated (see Hayes, 1979). At that point we say this device (or animal or person) has beliefs about heat and about this very room, and so forth, not only because of the system’s actual location in, and operations on, the world, but because we cannot imagine another niche in which it could be placed where it would work (see also Dennett, 1982, 1987a).

我们最初的简单恒温器有一种状态，我们称之为对特定锅炉的信念，即它是开着还是关着。为什么是那个锅炉呢？那么，你想说它是关于哪个锅炉的呢？信念是关于锅炉的，因为它是固定在锅炉上的。5^考虑到与世界实际存在的因果联系，即使是最小的因果联系，我们可以赋予设备的状态（某种意义上的意义）和真值条件，但用不同的最小联系来代替并完全改变内部状态的意义（在这种贫乏的意义上）是太容易了。但随着系统在感知上变得更加丰富，在行为上变得更加灵活，在不改变系统本身的组织的情况下，对系统与世界的实际联系进行替换变得越来越困难。如果你改变它的环境，它实际上会注意到，并作为回应改变它的内部状态。设备和环境之间出现了一种双向约束，即特异性越来越强。将设备固定在任何一种状态，它都需要一个非常特定的环境才能正常运行（您无法再轻松地将其从调节温度切换到调节速度或其他任何状态）；但同时，如果您不固定它所处的状态，而是将它放在一个变化的环境中，它的感官附件将足够敏感和有辨别力，能够对变化做出适当的反应，使系统进入新状态，在新环境中有效运行。有一种熟悉的方式来暗示系统组织与其环境之间存在的这种紧密关系：你可以说有机体不断地反映环境，或者说系统组织中存在环境的表征（或隐含在环境的表征中）。

Our original simple thermostat had a state we called a belief about a particular boiler, to the effect that it was on or off. Why about that boiler? Well, what other boiler would you want to say it was about? The belief is about the boiler because it is fastened to the boiler.⁵ Given the actual, if minimal, causal link to the world that happened to be in effect, we could endow a state of the device with meaning (of a sort) and truth conditions, but it was altogether too easy to substitute a different minimal link and completely change the meaning (in this impoverished sense) of that internal state. But as systems become perceptually richer and behaviorally more versatile, it becomes harder and harder to make substitutions in the actual links of the system to the world without changing the organization of the system itself. If you change its environment, it will notice, in effect, and make a change in its internal state in response. There comes to be a two-way constraint of growing specificity between the device and the environment. Fix the device in any one state and it demands a very specific environment in which to operate properly (you can no longer switch it easily from regulating temperature to regulating speed or anything else); but at the same time, if you do not fix the state it is in, but just plunk it down in a changed environment, its sensory attachments will be sensitive and discriminative enough to respond appropriately to the change, driving the system into a new state, in which it will operate effectively in the new environment. There is a familiar way of alluding to this tight relationship that can exist between the organization of a system and its environment: you say that the organism continuously mirrors the environment, or that there is a representation of the environment in—or implicit in—the organization of the system.

这并不是说我们只将信念和欲望归因于（或应该归因于）那些我们发现了内部表征的事物，而是当我们发现某个对象适合采用意向策略时，我们会努力将其某些内部状态或过程解释为内部表征。事物的某些内部特征之所以成为表征，只能是因为它调节意向系统的行为所起的作用。

It is not that we attribute (or should attribute) beliefs and desires only to things in which we find internal representations, but rather that, when we discover some object for which the intentional strategy works, we endeavor to interpret some of its internal states or processes as internal representations. What makes some internal feature of a thing a representation could only be its role in regulating the behavior of an intentional system.

现在，强调我们与恒温器的亲缘关系的理由应该很清楚了。从一个简单的恒温器过渡到一个真正具有周围世界内部表征的系统，并没有神奇的时刻。恒温器对世界的表征要求最低，更高级的恒温器对世界的表征要求更高，更高级的居家机器人对世界的表征要求更高。最后你来到了我们。我们与世界的联系是如此的多样和错综复杂，几乎不可能有任何替代——尽管在思想实验中可以清楚地想象到。希拉里·普特南想象了孪生地球，它和地球一模一样，甚至和你邻居的孪生地球复制品鞋子上的磨损痕迹一样，但它与地球在某些属性上有所不同，而这些属性完全超出了你的辨别能力的阈值。 （孪生地球上的水的化学分析结果不同。）如果你被瞬间带到孪生地球，换成孪生地球的复制品，你永远不会更聪明——就像简单的控制系统无法分辨它是在调节温度、速度还是水箱中的水量。对于像恒温器这样简单且缺乏感官的东西，设计出完全不同的孪生地球很容易，但你的内部组织对替代品的要求要严格得多。你的孪生地球和地球必须是虚拟复制品，否则你到达时的状态将发生巨大变化。

Now the reason for stressing our kinship with the thermostat should be clear. There is no magic moment in the transition from a simple thermostat to a system that really has an internal representation of the world around it. The thermostat has a minimally demanding representation of the world, fancier thermostats have more demanding representations of the world, fancier robots for helping around the house would have still more demanding representations of the world. Finally you reach us. We are so multifariously and intricately connected to the world that almost no substitution is possible—though it is clearly imaginable in a thought experiment. Hilary Putnam imagines the planet Twin Earth, which is just like Earth right down to the scuff marks on the shoes of the Twin Earth replica of your neighbor, but which differs from Earth in some property that is entirely beneath the thresholds of your capacities to discriminate. (What they call water on Twin Earth has a different chemical analysis.) Were you to be whisked instantaneously to Twin Earth and exchanged for your Twin Earth replica, you would never be the wiser—just like the simple control system that cannot tell whether it is regulating temperature, speed, or volume of water in a tank. It is easy to devise radically different Twin Earths for something as simple and sensorily deprived as a thermostat, but your internal organization puts a much more stringent demand on substitution. Your Twin Earth and Earth must be virtual replicas or you will change state dramatically on arrival.

那么，当您认为锅炉开着时，您的信念是关于哪个锅炉的呢？为什么是您地下室里的锅炉（而不是孪生地球上的孪生锅炉）。您的信念是关于其他哪个锅炉的呢？完成对您的信念的语义解释，确定您的信念的指称，就像恒温器的情况一样，需要有关您在世界上的实际嵌入的事实。当我们将信念归因于人时发现的解释原则和问题，与我们在研究将信念归因于恒温器这一荒谬但又非常简单的问题时发现的原则和问题相同。差异程度不同，但程度如此之大，以至于理解简单意向系统的内部组织，对于理解复杂意向系统（如人类）的内部组织来说，几乎没有什么基础。

So which boiler are your beliefs about when you believe the boiler is on? Why, the boiler in your cellar (rather than its twin on Twin Earth, for instance). What other boiler would your beliefs be about? The completion of the semantic interpretation of your beliefs, fixing the referents of your beliefs, requires, as in the case of the thermostat, facts about your actual embedding in the world. The principles, and problems, of interpretation that we discover when we attribute beliefs to people are the same principles and problems we discover when we look at the ludicrous, but blessedly simple, problem of attributing beliefs to a thermostat. The differences are of degree, but nevertheless of such great degree that understanding the internal organization of a simple intentional system gives one very little basis for understanding the internal organization of a complex intentional system, such as a human being.

11.3 为何有意策略有效？

11.3  Why does the intentional strategy work?

当我们转向这个问题：为什么意向策略如此有效时，我们发现这个问题很模糊，可以有两种截然不同的答案。如果意向系统是一个简单的恒温器，那么答案很简单：意向策略之所以有效，是因为恒温器设计精良；它被设计成一个可以轻松可靠地理解和操纵的系统。这是真的，但如果我们想要的是解释其性能的实际设计特征，那么答案就不是那么有用了。然而幸运的是，对于一个简单的恒温器来说，这些特征很容易发现和理解，因此我们的“为什么”问题的另一个答案很容易得到，而这实际上是关于机器如何工作的答案。

When we turn to the question of why the intentional strategy works as well as it does, we find that the question is ambiguous, admitting of two very different sorts of answer. If the intentional system is a simple thermostat, one answer is simply this: the intentional strategy works because the thermostat is well designed; it was designed to be a system that could be easily and reliably comprehended and manipulated from this stance. That is true, but not very informative, if what we are after are the actual features of its design that explain its performance. Fortunately, however, in the case of a simple thermostat those features are easily discovered and understood, so the other answer to our why question, which is really an answer about how the machinery works, is readily available.

如果所讨论的意向系统是人，那么我们的问题也存在歧义。对于意向策略为何有效这个问题，第一个答案是，进化将人类设计成理性的，相信他们应该相信的，想要他们应该想要的。我们是漫长而艰难的进化过程的产物，这一事实保证了在我们身上使用意向策略是万无一失的。这个答案具有真实和简洁的优点，但也令人惊讶地缺乏信息量。这个问题更难的版本实际上是在问大自然为我们提供的机器是如何运作的。而我们目前还不能对这个问题给出一个很好的答案。我们就是不知道。我们确实知道这个策略是如何运作的，我们知道它为什么有效的问题的简单答案，但知道这些对我们对困难答案没有多大帮助。

If the intentional system in question is a person, there is also an ambiguity in our question. The first answer to the question of why the intentional strategy works is that evolution has designed human beings to be rational, to believe what they ought to believe and want what they ought to want. The fact that we are products of a long and demanding evolutionary process guarantees that using the intentional strategy on us is a safe bet. This answer has the virtues of truth and brevity, but it is also strikingly uninformative. The more difficult version of the question asks, in effect, how the machinery which Nature has provided us works. And we cannot yet give a good answer to that question. We just do not know. We do know how the strategy works, and we know the easy answer to the question of why it works, but knowing these does not help us much with the hard answer.

然而，这并不是说缺乏理论。例如，斯金纳行为主义者会说，该策略之所以有效，是因为它对信念和欲望的归因实际上是对先前反应和强化历史的影响的难以想象的复杂描述的简写。说某人想要一些冰淇淋，就是说过去冰淇淋的摄入因结果而得到强化，在某些背景条件下（也太复杂而无法描述）产生了一种参与冰淇淋获取行为的倾向。在没有详细了解这些历史事实的情况下，我们仍然可以基于归纳理由做出精明的猜测；这些猜测体现在我们的意向立场主张中。即使所有这些都是真的，它也几乎不能告诉我们这些倾向是如何受到内部机制调节的。

It is not that there is any dearth of doctrine, however. A Skinnerian behaviorist, for instance, would say that the strategy works because its imputations of beliefs and desires are shorthand, in effect, for as yet unimaginably complex descriptions of the effects of prior histories of response and reinforcement. To say that someone wants some ice cream is to say that in the past the ingestion of ice cream has been reinforced in him by the results, creating a propensity under certain background conditions (also too complex to describe) to engage in icecream-acquiring behavior. In the absence of detailed knowledge of those historical facts we can nevertheless make shrewd guesses on inductive grounds; these guesses are embodied in our intentional stance claims. Even if all this were true, it would tell us very little about the way such propensities were regulated by the internal machinery.

目前更流行的解释是，策略如何运作和机制如何运作的解释将（大致）一致：对于每个可预测归因的信念，都会有一个功能上显著的机器内部状态，可以分解成功能部分，就像表达信念的句子可以分解成部分（即单词或术语）一样。我们归因于理性生物的推论将通过硬件中的物理因果过程反映出来；所相信命题的逻辑形式将以状态的结构形式复制与他们相对应。这是假设我们的大脑中存在一种思维语言，我们的大脑最终将被理解为符号操作系统，至少与计算机大致类似。在名为认知科学的新研究计划中，目前正在探索这种观点的许多不同版本，只要人们允许对基本大胆的主张进行很大的削弱，我认为它的某个版本将被证明是正确的。

A currently more popular explanation is that the account of how the strategy works and the account of how the mechanism works will (roughly) coincide: for each predictively attributable belief, there will be a functionally salient internal state of the machinery, decomposable into functional parts in just about the same way the sentence expressing the belief is decomposable into parts—that is, words or terms. The inferences we attribute to rational creatures will be mirrored by physical, causal processes in the hardware; the logical form of the propositions believed will be copied in the structural form of the states in correspondence with them. This is the hypothesis that there is a language of thought coded in our brains, and our brains will eventually be understood as symbol manipulating systems in at least rough analogy with computers. Many different versions of this view are currently being explored, in the new research program called cognitive science, and provided one allows great latitude for attenuation of the basic, bold claim, I think some version of it will prove correct.

但我不认为这是显而易见的。那些认为这种理论显然或不可避免地会被证明是正确的人（很多人都这么认为），混淆了两个经验主张。第一个是，意向立场描述产生了一个客观的、真实的世界模式——我们想象中的火星人所错过的模式。这是一个经验主张，但却是毫无疑问的。第二个是，这种真实模式是由智慧生物大脑中另一种与它大致同构的真实模式产生的。怀疑第二种真实模式的存在并不是怀疑第一种模式的存在。有理由相信第二种模式，但这些理由并不是压倒性的。我能给出的关于这些理由的最好的简单说明如下。

But I do not believe that this is obvious. Those who think that it is obvious, or inevitable, that such a theory will prove true (and there are many who do), are confusing two empirical claims. The first is that intentional stance description yields an objective, real pattern in the world—the pattern our imaginary Martians missed. This is an empirical claim, but one that is confirmed beyond skepticism. The second is that this real pattern is produced by another real pattern roughly isomorphic to it within the brains of intelligent creatures. Doubting the existence of the second real pattern is not doubting the existence of the first. There are reasons for believing in the second pattern, but they are not overwhelming. The best simple account I can give of the reasons is as follows.

随着我们从简单的恒温器、复杂的机器人到人类的复杂程度不断上升，我们发现，我们为设计具有必要行为的系统所做的努力越来越与组合爆炸问题相冲突。将某个参数增加 10%（增加 10% 的输入、增加要控制的行为的自由度、增加要识别的单词等）往往会使所设计系统的内部复杂性增加几个数量级。事情很快就会失控，例如，可能导致计算机程序淹没最大、最快的机器。现在，大脑不知何故解决了组合爆炸问题。它是一个由数十亿个细胞组成的庞大网络，但仍然是有限的、紧凑的、可靠的和快速的，并且能够几乎无限制地学习新的行为、词汇和理论。一些优雅的、生成的、可无限扩展的表示原理一定是造成这种情况的原因。我们只有一个这样的表示系统模型：人类语言。因此，关于思维语言的争论可以归结为：它还能是什么？迄今为止，我们无法详细想象任何合理的替代方案。我认为，这是一个很好的理由，可以作为科学策略建议我们尽可能地以各种形式追求假设。6但是^，如果我们牢记其必然正确性远未得到保证，我们将更加谨慎、富有成效地进行这种探索。只要人们误以为经验假设必然正确，就无法很好地理解它。

As we ascend the scale of complexity form simple thermostat, through sophisticated robot, to human being, we discover that our efforts to design systems with the requisite behavior increasingly run foul of the problem of combinatorial explosion. Increasing some parameter by, say, ten percent—ten percent more inputs or more degrees of freedom in the behavior to be controlled or more words to be recognized or whatever—tends to increase the internal complexity of the system being designed by orders of magnitude. Things get out of hand very fast and, for instance, can lead to computer programs that will swamp the largest, fastest machines. Now somehow the brain has solved the problem of combinatorial explosion. It is a gigantic network of billions of cells, but still finite, compact, reliable, and swift, and capable of learning new behaviors, vocabularies, theories, almost without limit. Some elegant, generative, indefinitely extensible principles of representation must be responsible. We have only one model of such a representation system: a human language. So the argument for a language of thought comes down to this: what else could it be? We have so far been unable to imagine any plausible alternative in any detail. That is a good reason, I think, for recommending as a matter of scientific tactics that we pursue the hypothesis in its various forms as far as we can.⁶ But we will engage in that exploration more circumspectly, and fruitfully, if we bear in mind that its inevitable rightness is far from assured. One does not well understand even a true empirical hypothesis so long as one is under the misapprehension that it is necessarily true.

第 11 章注释

Notes for Chapter 11

1.   1.任何人的大多数信念都必定是真实的这一观点对某些人来说似乎是显而易见的。在奎因、普特南、舒梅克、戴维森和我本人的著作中都可以找到对这一观点的支持。其他人也同样觉得这个想法难以置信——所以很可能双方都把不同的现象称为信念。一旦人们区分了信念和观点（我的专业意义上的——丹尼特，1978b），根据这一区别，观点是受语言影响的、相对复杂的认知状态——大致是对某个特定的、公式化的句子的真实性下注的状态——人们就会发现大多数信念是真实的这一说法几乎是微不足道的。对一些外围问题的思考应该可以揭示这一点。以德谟克利特为例，他有一个系统的、包罗万象的，但（为了论证起见，我们可以说）完全是错误的物理学。他把所有的事情都搞错了，尽管他的观点成立并且具有某种系统的效用。但即使学术界允许我们归因于德谟克利特的每一项主张（无论是其著作中明确的还是隐含的）都是错误的，这些主张也只代表了他信念中微不足道的一部分，其中既包括他必然拥有的大量单调乏味的既定信念（关于他住在哪所房子里，要寻找什么样的好凉鞋，等等），也包括那些随着他的感知经验的变化而出现和消失的数百万偶然信念。

2.   1.  The idea that most of anyone’s beliefs must be true seems obvious to some people. Support for the idea can be found in works by Quine, Putnam, Shoemaker, Davidson, and myself. Other people find the idea equally incredible—so probably each side is calling a different phenomenon belief. Once one makes the distinction between belief and opinion (in my technical sense—Dennett, 1978b), according to which opinions are linguistically infected, relatively sophisticated cognitive states—roughly states of betting on the truth of a particular, formulated sentence—one can see the near triviality of the claim that most beliefs are true. A few reflections on peripheral matters should bring it out. Consider Democritus, who had a systematic, all-embracing, but (let us say, for the sake of argument) entirely false physics. He had things all wrong, though his views held together and had a sort of systematic utility. But even if every claim that scholarship permits us to attribute to Democritus (either explicit or implicit in his writings) is false, these represent a vanishingly small fraction of his beliefs, which include both the vast numbers of humdrum standing beliefs he must have had (about which house he lived in, what to look for in a good pair of sandals, and so forth) and also those occasional beliefs that came and went by the millions as his perceptual experience changed.

但是，有人可能会说，将德谟克利特的平淡无奇的信念与他的科学相隔离，依赖于观察真理和理论真理之间无法辩驳的区别；德谟克利特的所有信念都承载着理论，而由于他的理论是错误的，所以这些信念就是错误的。答复如下：既然所有观察信念都承载着理论，为什么我们要选择德谟克利特明确、复杂的理论（体现在他的《观点》中）作为他日常观察的理论呢？请注意，德谟克利特最不讲理论的同胞也拥有无数承载着理论的观察信念——而且在某种意义上，他并没有因此而变得更明智。为什么我们不认为德谟克利特的观察也承载着同样的（大概是无害的）理论呢？如果德谟克利特忘记了他的理论，或者改变了主意，他的观察信念基本不会受到影响。就他的复杂理论在他的日常行为和期望等方面发挥了明显作用而言，用复杂理论来表达他平淡无奇的信念是相当合适的，但这不会产生一个主要错误的信念目录，因为他的信念很少会受到影响。（然而，理论对观察的影响往往被低估。请参阅 Churchland (1979) 中关于理论和经验之间有时可能存在的紧密关系的戏剧性和令人信服的例子。）（本说明中的讨论是从与 Paul 和 Patricia Churchland 以及 Michael Stack 的一次有益对话中提炼出来的。）

But, it may be urged, this isolation of his humdrum beliefs from his science relies on an insupportable distinction between truths of observation and truths of theory; all Democritus’s beliefs are theory-laden, and since his theory is false, they are false. The reply is as follows: Granted that all observation beliefs are theory-laden, why should we choose Democritus’s explicit, sophisticated theory (couched in his opinions) as the theory with which to burden his quotidian observations? Note that the least theoretical compatriot of Democritus also had myriads of theory-laden observation beliefs—and was, in one sense, none the wiser for it. Why should we not suppose Democritus’s observations are laden with the same (presumably innocuous) theory? If Democritus forgot his theory, or changed his mind, his observational beliefs would be largely untouched. To the extent that his sophisticated theory played a discernible role in his routine behavior and expectations and so forth, it would be quite appropriate to couch his humdrum beliefs in terms of the sophisticated theory, but this will not yield a mainly false catalogue of beliefs, since so few of his beliefs will be affected. (The effect of theory on observation is nevertheless often underrated. See Churchland (1979) for dramatic and convincing examples of the tight relationship that can sometimes exist between theory and experience.) (The discussion in this note was distilled from a useful conversation with Paul and Patricia Churchland and Michael Stack.)

1.   2.牛津的一位听众指出，如果火星人将地球人纳入其物理立场范围（我并没有明确排除这种可能性），他就不会对地球人的预测感到惊讶。他确实会准确预测出地球人说火星语时产生的 X 射线调制模式。没错，但当火星人写下他的计算结果时，他对地球人预测的预测就会一字不差地出现，就像在通灵板上一样，而让火星人感到吃惊的是，这个机制，这个装扮成火星人的地球人预测器，是如何在与火星人需要了解的事件信息如此隔绝的情况下，得出火星人这句真正的句子的，以便对汽车的到来做出自己的预测。

2.   2.  A member of the audience in Oxford pointed out that if the Martian included the Earthling in his physical stance purview (a possibility I had not explicitly excluded), he would not be surprised by the Earthling’s prediction. He would indeed have predicted exactly the pattern of X-ray modulations produced by the Earthling speaking Martian. True, but as the Martian wrote down the results of his calculations, his prediction of the Earthling’s prediction would appear, word by Martian word, as on a Ouija board, and what would be baffling to the Martian was how this chunk of mechanism, the Earthling predictor dressed up like a Martian, was able to yield this true sentence of Martian when it was so informationally isolated from the events the Martian needed to know of in order to make his own prediction about the arriving automobile.

3.   3.可能存在不需要交流、预测、观察等技能的智慧生物吗？可能存在奇妙、灵巧、无懈可击的生物，它们不具备这些行为方式，但我看不出我们为什么称它们为智慧生物。

4.   3.  Might there not be intelligent beings who had no use for communicating, predicting, observing,…? There might be marvelous, nifty, invulnerable entities lacking these modes of action, but I cannot see what would lead us to call them intelligent.

5.   4.约翰·麦卡锡 (John McCarthy) 的密码学类比很好地说明了这一点。密文语料库越大，出现双重、系统上不相关的解密的可能性就越小。有关应用于机器（明确包括恒温器）的意向立场的原则和前提的非常有用的讨论，请参阅 McCarthy，1979 年。

6.   4.  John McCarthy’s analogy to cryptography nicely makes this point. The larger the corpus of cipher text, the less chance there is of dual, systematically unrelated decipherings. For a very useful discussion of the principles and presuppositions of the intentional stance applied to machines—explicitly including thermostats—see McCarthy, 1979.

7.   5。  这一思想实际上是归入“论物信仰”范畴的各种不同思想的祖先。如果人们从这一思想出发，去探究它的后代，就能更好地看到它们的困难，以及如何解决它们。（有关这一主题的更多信息，请参阅 Dennett，1982 年。）

8.   5.  This idea is the ancestor in effect of the species of different ideas lumped together under the rubric of de re belief. If one builds from this idea toward its scions, one can see better the difficulties with them, and how to repair them. (For more on this topic, see Dennett, 1982.)

9.   6.迄今为止提出的所有心理表征的思维语言模型都以这样或那样的方式成为组合爆炸的牺牲品，这一事实应该会抑制人们参与福多尔恰当地称之为“城里唯一的游戏”的热情。

10.   6.  The fact that all language-of-thought models of mental representation so far proposed fall victim to combinatorial explosion in one way or another should temper one’s enthusiasm for engaging in what Fodor aptly calls “the only game in town”.

第 12 章注释

Notes for Chapter 12

1.   1.当然，我并不是说 Schank 本人也坚持这些主张。

2.   1.  I am not saying, of course, that Schank himself is committed to these claims.

3.   2.此外，“理解”既意味着拥有心理（意向）状态，也意味着这些状态的真实性（有效性、成功性）。为了讨论的目的，我们只关注这些状态的拥有。

4.   2.  Also, “understanding” implies both the possession of mental (intentional) states and the truth (validity, success) of these states. For the purposes of this discussion, we are concerned only with the possession of the states.

5.   3.意向性的定义是某些心理状态的特征，这些心理状态针对或涉及世界上的物体和事态。因此，信念、欲望和意图是意向性状态；无指向的焦虑和抑郁则不是。（有关进一步讨论，请参阅 Searle，1979 年）。

6.   3.  Intentionality is by definition that feature of certain mental states by which they are directed at or are about objects and states of affairs in the world. Thus, beliefs, desires, and intentions are intentional states; undirected forms of anxiety and depression are not. (For further discussion, see Searle, 1979).

14.1 简介

14.1  Introduction

自 20 世纪 60 年代行为主义消亡以来，计算主义一直是认知心理学的主导思想。计算理论将人类认知过程视为一种信息处理，将实现此类处理的系统视为符号操作系统。将设备描述为符号操纵器意味着可以将其某些内部状态视为特定领域中的属性或对象的表示。例如，视觉计算理论假设内部状态可以解释为表示远端场景的深度。

The dominant program in cognitive psychology since the demise of behaviorism in the 1960s has been computationalism. Computational theories treat human cognitive processes as a species of information processing, and the systems that implement such processing as symbol-manipulating systems. Describing a device as a symbol manipulator implies that it is possible to treat some of its internal states as representations of properties or objects in a particular domain. Computational theories of vision, for example, posit internal states that can be interpreted as representing the depth of the distal scene.

关于计算理论所假设的状态所赋予的表征内容的性质和功能，存在着相当大的分歧。人们普遍认为，这类理论尊重杰里·福多尔 (Jerry Fodor) (1980) 所说的“形式性条件”，该条件要求计算过程只能访问定义它们所基于的表征的形式属性（即非语义属性）。通过尊重形式性条件，计算主义有望回答心灵哲学中最紧迫的问题之一——表征心理状态如何在行为的产生中产生因果效应？根据计算主义的说法，表征心理状态由于（大致）其结构属性而具有因果作用。1^但这种优势是有代价的。计算描述的形式特征似乎没有为它所表征的心理状态的语义属性留下任何实际工作。因此，有些人认为计算主义支持一种消除主义，即否认有意描述的状态在心理学中发挥真正的解释作用的论点（例如，参见 Stich，1983）。如果计算状态的内容确实没有解释作用，那么计算心理学所表征的心理状态与我们的常识解释实践所表征的心理状态之间的关系（确实涉及内容）就相当模糊了。

There has been considerable disagreement about the nature and function of representational contents assigned to the states posited by computational theories. It is widely thought that such theories respect what Jerry Fodor (1980) has called the “formality condition,” which requires that computational processes have access only to the formal (that is, nonsemantic) properties of the representations over which they are defined. It is by respecting the formality condition that computationalism promises to answer one of the most pressing problems in the philosophy of mind—how can representational mental states be causally efficacious in the production of behavior? Representational mental states, according to computationalism, have their causal roles in virtue of (roughly) their structural properties.¹ But this advantage comes at a price. The formal character of computational description appears to leave no real work for the semantic properties of the mental states it characterizes. Thus, computationalism has been thought by some to support a form of eliminativism, the thesis that denies that intentionally characterized states play a genuinely explanatory role in psychology (see, for example, Stich, 1983). If the content of computational states is indeed explanatorily idle, then the relation between psychological states, as characterized by computational psychology, and psychological states as characterized by our commonsense explanatory practices, which do advert to content, is quite obscure.

在本文中，我阐明并捍卫了一种策略，用于协调计算描述的形式特征与对心理解释有用性的承诺。内容。我认为内容在计算理论中并不起个体或分类作用——过程的计算表征是一种形式表征。尽管如此，内容在认知能力的计算描述中确实发挥着真正的解释作用。内容归因将内部过程的形式表征与主体的环境联系起来，使计算理论能够解释过程的运行如何构成该环境中认知能力的运用。我通过参考 David Marr 的早期视觉理论²来支持我对内容在计算心理学中的作用的解释，部分原因是它受到了哲学家的极大关注；然而，我的论点取决于计算方法的一般特征，因此适用于一般的计算理论。

In this paper I articulate and defend a strategy for reconciling the formal character of computational description with a commitment to the explanatory usefulness of mental content. I argue that content does not play an individuative or taxonomic role in computational theories—a computational characterization of a process is a formal characterization. Nonetheless, content does play a genuine explanatory role in computational accounts of cognitive capacities. Content ascriptions connect the formal characterization of an internal process with the subject’s environment, enabling the computational theory to explain how the operation of the process constitutes the exercise of a cognitive capacity in that environment. I support my account of the role of content in computational psychology by reference to David Marr’s theory of early vision,² in part because it has received a great deal of attention from philosophers; however, my argument depends on general features of computational methodology, and so applies to computational theories generally.

最近，为了调和计算和内容，人们开始诉诸狭义内容的概念，即内容依赖于主体的内在物理状态。狭义内容的支持者迄今为止未能阐明一个明确适用于心理学中真正的解释工作的概念。3我认为^，在计算解释中起核心作用的通常是广义内容，尽管我确实确定了狭义内容归因所发挥的特定（且有限的）作用。

Recent attempts to reconcile computation and content have appealed to a notion of narrow content, that is, content that supervenes on intrinsic physical states of the subject. Proponents of narrow content have so far failed to articulate a notion that is clearly suitable for genuine explanatory work in psychology.³ I argue that it is typically broad content that plays a central role in computational explanation, though I do identify a specific (and limited) function served by narrow content ascription.

14.2 为什么计算理论不是有意的

14.2  Why Computational Theories Are Not Intentional

有人可能会说，尽管有形式条件，但计算理论在以下意义上是有意的：它们假设的状态不仅具有表征内容，而且它们所具有的内容在理论中发挥着个体作用。换句话说，计算理论根据状态的内容对状态进行分类。

It might be argued that, the formality condition notwithstanding, computational theories are intentional in the following sense: The states they posit not only have representational content, but the content they have plays an individuative role in the theory. In other words, computational theories taxonomize states by reference to their contents.

声称计算认知理论在上述意义上是有意的，其动机并不难理解。请考虑以下段落：

The motivation for the claim that computational theories of cognition are intentional in the above sense is not hard to understand. Consider the following passages:

这两段文字背后的论点可以粗略地表达如下：

The argument underlying both passages can be expressed somewhat crudely as follows:

这一论点的基础是直觉，即科学解释应该“匹配”（在某种意义上）其被解释项。威尔逊赞同这种约束，他称之为理论适当性。

Underlying the argument is the intuition that scientific explanations should “match” (in some sense) their explananda. Wilson endorses a constraint of this sort, which he calls theoretical appropriateness.

“解释层次”的概念有些模糊，但为了论证的目的，我们假设存在一个独特的解释层次，即所有且只有该层次的解释才涉及内容的归属。如果理论的适当性是科学解释的必要条件，那么对有意表征现象的解释本身就应该关注有意表征的状态^。4

The notion of a “level of explanation” is somewhat vague, but let us assume, for the sake of argument, that there is a unique level of explanation such that all and only explanations at that level involve ascriptions of content. If theoretical appropriateness is a desideratum of scientific explanation, then an explanation of intentionally characterized phenomena should itself advert to intentionally characterized states.⁴

但是，当我们考虑计算解释时，一个尚未解决的紧张关系浮现出来。计算理论的解释项是故意指定的，这一事实表明，计算状态本质上是通过参考其内容而个体化的。如果计算理论不是有意的，那么计算理论如何解释有意表征的现象？但是形式性条件施加了相反的压力。在要求计算过程只能访问它们所定义的表征状态的非语义属性时，它表明计算个体化是非语义的，或者用 Fodor 的术语来说，是形式化的。计算分类法是有意的还是形式化的？此时，转向一个完善的例子会有所帮助。

An unresolved tension surfaces, though, when we consider computational explanation. The fact that the explananda of computational theories are intentionally specified suggests that computational states are essentially individuated by reference to their contents. If computational theories are not intentional, then how can computational theories explain intentionally characterized phenomena? But the formality condition exerts an opposite pressure. In requiring that computational processes have access only to the nonsemantic properties of the representational states over which they are defined, it suggests that computational individuation is nonsemantic, or in Fodor’s terminology, formal. Are computational taxonomies intentional or formal? At this point it is helpful to turn to a well-developed example.

Marr 视觉理论的解释者认为，视觉状态在该理论中根据其内容而个体化，因此该理论是意向性的（参见 Burge，1986；Kitcher，1988；Segal，1989、1991；Davies，1991；Morton，1993；Shapiro，1993）。尽管关于 Marrian 结构的内容类型（广义或狭义）存在很多分歧，但内容在该理论中发挥个体作用的假设被认为不需要明确的论证。5 ^Burge表示，从“该理论的顶层明确地以意向性术语表述”这一事实来看，该理论是意向性的“证据充分”。我认为，在将内容解释为个体性时，Marr 的解释者误解了内容在计算理论中的作用。

Interpreters of Marr’s theory of vision have assumed that visual states are individuated in the theory by reference to their contents, hence that the theory is intentional (see Burge, 1986; Kitcher, 1988; Segal, 1989, 1991; Davies, 1991; Morton, 1993; Shapiro, 1993). Although there has been a good deal of disagreement about the sort of content (broad or narrow) that Marrian structures have, the assumption that content plays an individuative role in the theory has not been thought to require explicit argument.⁵ Burge says that it is “sufficiently evident” that the theory is intentional from the fact that “the top levels of the theory are explicitly formulated in intentional terms” (1986, 55). I shall argue that in construing content as individuative, interpreters of Marr have misconstrued the role of content in computational theories.

虽然 Marr 在非正式阐述各种视觉过程时，通常通过参考远端场景的特征来描述它们，但不应过分解读这一事实。这些过程也具有正式的特征。它们必须是正式的——Marr 的视觉理论是一种计算理论，正式的特征保证了它们是可编程的（因此，是物理上可实现的）。问题是，哪种特征可以完成个体工作？

While it is true that in his informal exposition of the various visual processes Marr typically characterizes them by reference to features of the distal scene, one should not read too much into this fact. The processes are also characterized formally. They have to be—Marr’s theory of vision is a computational theory, and a formal characterization guarantees that they are programmable (hence, physically realizable). The question is, which characterization does the individuative work?

Marr 有力地论证了信息处理系统应该在三个不同的描述层次上进行分析。Marr 称之为“顶层”理论，计算，是系统^计算的功能的特征——系统做什么。算法层指定了计算功能的算法，实现层描述了该过程的物理实现方式。6 Marr 层次结构的顶层有时与 Pylyshyn 的语义层（Pylyshyn，1984）和 Newell 的知识层（Newell，1982）相一致。换句话说，计算理论本质上被解释为一种机制的意向性或语义性表征。但这样的解释让 Marr 坚持认为算法的探索必须等待计算理论的精确规范这一点有些令人费解。他说，“除非正确地制定了过程的计算理论，否则算法几乎肯定是错误的”（1982，124），这表明顶层应该理解为提供设备的功能理论表征。事实上，Marr 明确指出，计算理论是对各种处理模块计算出的函数的数学表征。在描述表征图像初始滤波（计算图像的拉普拉斯算子与高斯卷积）的数学公式时，Marr 说道：

Marr argued persuasively that an information-processing system should be analyzed at three distinct levels of description. The “top” level, which Marr called the theory of the computation, is a characterization of the function computed by the system—what the system does. The algorithmic level specifies an algorithm for computing the function, and the implementation level describes how the process is realized physically.⁶ The top level in Marr’s hierarchy is sometimes identified with Pylyshyn’s semantic level (Pylyshyn, 1984) and Newell’s knowledge level (Newell, 1982). In other words, the theory of the computation has been construed as essentially an intentional or semantic characterization of a mechanism. But such a construal makes somewhat puzzling Marr’s insistence that the search for the algorithm must await the precise specification of the theory of the computation. He says, “unless the computational theory of the process is correctly formulated, the algorithm will almost certainly be wrong” (1982, 124), suggesting that the top level should be understood to provide a function-theoretic characterization of the device. Indeed, Marr explicitly points out that the theory of the computation is a mathematical characterization of the function(s) computed by the various processing modules. In describing the mathematical formula that characterizes the initial filtering of the image (the calculation of the Laplacian of the image convolved with a Gaussian), Marr says the following:

∇ ² G是一个函数，它以二维强度数组I ( x, y ) 作为参数，以数组中各点 ( x, y )处强度的各向同性变化率作为值。此函数的实现用于 Marr 和 Hildreth (1980) 的边缘检测模型，以检测零交叉点。（零交叉点是函数值改变符号的点。零交叉对应于图像中的急剧强度变化。）Marr 承认，视网膜计算的函数的数学规范可能无法使视网膜所做的事情变得清晰。尽管如此，从信息处理的角度来看，正式规范是“充分的”。更准确地说，它是算法的正确规范所关键依赖的描述。

∇² G is a function that takes as arguments two-dimensional intensity arrays I(x, y) and has as values the isotropic rates of change of intensity at points (x, y) in the array. The implementation of this function is used in Marr and Hildreth’s (1980) model of edge detection to detect zero-crossings. (A zero-crossing is a point where the value of a function changes its sign. Zero-crossings correspond to sharp intensity changes in the image.) Marr grants that the mathematical specification of the function computed by the retina may not make what the retina does perspicuous. Nonetheless, from an information-processing point of view, the formal specification is “adequate.” More precisely, it is the description upon which the correct specification of the algorithm crucially depends.

顶层提供数学表征的说法并不意味着说视觉系统将光强度值的表示作为输入并产生形状的表示作为输出是错误的。我并不否认计算过程具有真正的意向（语义）描述。出于某些目的，正如我们将在下一节中看到的那样，过程的意向描述将比形式表征更可取。说机制计算的功能的意向表征位于 Marr 层次结构的顶层并没错，尽管意向表征提供了设备功能的外在描述，并且没有^将计算过程个体化。为了个体化的目的，计算理论给出的精确数学描述才是重要的描述^。7、8

The claim that the top level provides a mathematical characterization does not imply that it is wrong to speak of the visual system as taking representations of light intensity values as input and yielding representations of shape as output. I am not denying that computational processes have true intentional (semantic) descriptions. For some purposes, as we shall see in the next section, an intentional description of a process will be preferable to a formal characterization. It is not incorrect to say that an intentional characterization of the function computed by a mechanism resides at the top level in Marr’s hierarchy, although the intentional characterization provides an extrinsic description of what the device does, and does not individuate the computational process. For the purpose of individuation, the precise mathematical description given by the theory of the computation is the description that counts.⁷ ,⁸

如果正如我所论证的那样，计算说明的顶层提供了设备的纯数学特征，那么就不太可能将第二层或算法层解释为有意的。（据我所知，只有 Burge 将算法层解释为有意的，显然是因为在讨论各种可能的算法时，Marr 有时会使用有意的语言。）理论的算法层只是指定系统如何计算顶层以数学术语表征的函数。

If, as I have argued, the top level of a computational account provides a purely mathematical characterization of a device, then there is little temptation to construe the second, or algorithmic level, as intentional. (Only Burge, as far as I know, construes the algorithmic level as intentional, apparently because in discussing various possible algorithms Marr sometimes employs intentional language.) The algorithmic level of theory simply specifies how the function characterized in mathematical terms at the top level is computed by the system.

14.3 内容的解释作用

14.3  The Explanatory Role of Content

我曾论证过，马尔的视觉理论不是有意的。我的论点诉诸于计算方法论的一般特征；如果我是对的，那么认知的计算理论就不是有意的——这些理论所表征的状态和过程并不是根据赋予它们的表征内容而个体化的。形式化（即数学化）表征起到了分类作用。

I have argued that Marr’s theory of vision is not intentional. My argument appeals to general features of computational methodology; if I am right, then computational theories of cognition are not intentional—the states and processes characterized by such theories are not individuated by reference to the representational contents ascribed to them. The formal—namely mathematical—characterization does the taxonomic work.

让我们暂时考虑一下计算理论不是有意为之这一说法的含义。从计算的角度来看，使用相同算法计算相同数学函数的两种机制是相同的机制，即使它们可能部署在完全不同的环境中。计算描述是对机制的独立于环境的表征。9^由于计算过程通常被解释为模块化过程，因此即使是内部环境也与计算过程的类型个体化无关。想象视觉系统的一个组成部分，称为 visex ，它根据双眼视差信息计算出视觉场景深度的表示。10现在想象^一下，在某个实际或想象的生物的听觉系统中，有一个在物理上与 visex 相同的组成部分。将这个组成部分称为audex。根据适合这种生物的听觉处理理论，audex 计算出某些声音特性的表示。我们可以想象将特定的 visex 和 audex 从视觉和听觉系统中的正常嵌入中移除并进行切换。由于这两个组件在假设上在物理上是相同的，因此它们计算同一类函数。切换不会对生物的行为或它们头脑中发生的事情产生明显的影响。尽管它们的正常内部环境存在差异，但这两种机制在计算上是相同的。

Let us consider for a moment the implications of the claim that computational theories are not intentional. Two mechanisms that compute the same mathematical function, using the same algorithm, are, from a computational point of view, the same mechanism, even though they may be deployed in quite different environments. A computational description is an environment-independent characterization of a mechanism.⁹ Inasmuch as computational processes are generally construed as modular processes, even the internal environment is irrelevant to the type-individuation of a computational process. Imagine a component of the visual system, called the visex, that computes a representation of the depth of the visual scene from information about binocular disparity.¹⁰ Now imagine that within the auditory system of some actual or imagined creature there is a component that is physically identical to the visex. Call this component the audex. According to the theory of auditory processing appropriate to this creature, the audex computes a representation of certain sonic properties. We can imagine a particular visex and audex removed from their normal embeddings in visual and auditory systems respectively and switched. Since the two components are by hypothesis physically identical, they compute the same class of functions. The switch will make no discernible difference to the behavior of the creatures, nor to what is going on inside their heads. The two mechanisms are computationally identical, despite the difference in their normal internal environments.

也许应该注意到，描述 visex 的视觉理论将其描述为根据视差计算深度表示，而不是计算某些声音属性的表示，尽管如果将其嵌入到不同的内部环境中，它将执行后者。重要的一点是，假设的结构没有独立于它们通常所处的环境（内部和外部）的内容。这就是计算过程的有意描述是外在描述的意义。原始草图中的结构包含来自几个不同的 ∇ ² G通道的信息，并为 Marr 理论所描述的大多数模块化过程提供输入，这些结构与诸如物体边界或照明变化等显着的远端属性可靠地相关，并被 Marr 描述为代表这些属性。然而，在某些完全不同的环境中，相同的结构可能与不同的远端属性相关，或者可能与世界的客观特征无关。在后一个世界中，这些结构不会代表任何东西，除了图像的特征。在那个世界中他们不会有任何远端内容。

It will perhaps be noted that the visual theory that describes the visex characterizes it as computing a representation of depth from disparity, and not as computing a representation of certain sonic properties, although it would do the latter if it were embedded in a different internal environment. The important point is that the postulated structures have no content considered independently of the environment (internal and external) in which they are normally situated. This is the sense in which an intentional characterization of a computational process is an extrinsic description. Structures in the raw primal sketch, which contains information from several distinct ∇²G channels and provides the input to most of the modular processes characterized by Marr’s theory, are reliably correlated with such salient distal properties as object boundaries or changes in illumination, and are described by Marr as representing these properties. In some radically different environment, however, the same structures may be correlated with different distal properties, or perhaps with no objective feature of the world. In the latter world, the structures would not represent anything, except perhaps features of the image. They would have no distal content in that world.

我想强调的一点是，计算机制的刻意表征隐含着与该机制通常所处环境的相对关系。计算理论提供的数学表征则不然。只有数学表征才能选出计算机制的基本属性。刻意表征并非必不可少，因为在某些可能的情况下它并不适用。

The point I wish to underscore is that an intentional characterization of a computational mechanism involves an implicit relativization to the context in which the mechanism is normally embedded. The mathematical characterization provided by the theory of the computation does not. Only the mathematical characterization picks out an essential property of a computational mechanism. The intentional characterization is not essential, since in some possible circumstances it would not apply.

那么，表征内容在认知过程的计算描述中扮演什么角色呢？如果不是从本质上描述认知过程的话。我曾在其他地方论证过，Egan (1992) 认为，语义解释在计算心理学中扮演的角色类似于解释模型在物理科学中扮演的角色。这种说法在两个方面是正确的。首先，有意描述一个本质上形式化的过程具有解释性功能，可以解释形式化描述，而形式化描述本身可能并不清晰。其次，当一个理论没有完全说明时（如 Marr 的理论），对该理论模型的研究通常有助于随后对该理论本身的阐述。计算理论家可能会部分地参考某些表征域的特征来描述计算，希望以后能提供形式化的细节（即理论）。

What, then, is the role that representational content plays in computational accounts of cognitive processes, if not to essentially characterize cognitive processes? I have argued elsewhere Egan (1992) that semantic interpretations play a role in computational psychology analogous to the role played by explanatory models in the physical sciences. There are two senses in which this is true. In the first place, an intentional characterization of an essentially formal process serves an expository function, explicating the formal account, which might not itself be perspicuous. Secondly, when a theory is incompletely specified (as is Marr’s theory), the study of a model of the theory can often aid in the subsequent elaboration of the theory itself. A computational theorist may resort to characterizing a computation partly by reference to features of some represented domain, hoping to supply the formal details (i.e., the theory) later.

虽然我认为与物理学中的模型的类比很有趣也很有用，但计算过程的有意解释所发挥的最重要的作用却是心理学所独有的。预先定义心理学理论领域的问题通常以有意的术语来表达。例如，我们希望视觉理论能够告诉我们视觉系统如何从二维图像中包含的信息中检测深度等。对假设的计算过程的有意规范表明，该理论确实回答了这些问题。只有在将系统的某些状态解释为远端属性（如深度或表面反射率）的表示的情况下，计算理论给出的数学特征的过程才会被揭示为视觉。因此，内容归因起着至关重要的解释作用：我们需要它们来解释正式表征的过程的运行如何构成在通常部署该过程的环境中认知能力的运用。

Though the analogy with models in physics is, I think, interesting and useful, the most important function served by intentional interpretations of computational processes is unique to psychology. The questions that antecedently define a psychological theory’s domain are usually couched in intentional terms. For example, we want a theory of vision to tell us, among other things, how the visual system can detect depth from information contained in two-dimensional images. An intentional specification of the postulated computational processes demonstrates that these questions are indeed answered by the theory. It is only under an interpretation of some of the states of the system as representations of distal properties (like depth, or surface reflectance) that the processes given a mathematical characterization by a computational theory are revealed as vision. Thus content ascriptions play a crucial explanatory role: we need them to explain how the operation of a formally characterized process constitutes the exercise of a cognitive capacity in the environment in which the process is normally deployed.

让我们回到之前关于计算认知理论是有意的这一论证：

Let us return for a moment to the argument considered earlier for the claim that computational theories of cognition are intentional:

这个论证的前提是正确的——定义心理学理论解释领域的问题通常用意向性术语来表达——但是，正如我们所见，这并不意味着该理论将其所描述的状态和过程必然描述为意向性的。计算状态和过程在独立于环境考虑时通常没有真正的意向性描述。因此，意向性表征不是计算理论的个体装置的一部分。从这个意义上说，(C) 是错误的。然而，这个论证确实包含一个重要的见解：需要意向性表征来将计算理论与其前理论的解释项联系起来。只有当理论以形式术语表征的状态被解释为远端属性的表示时，才能解释视觉系统如何从二维图像中包含的信息中检测场景的深度。

The premise of the argument is true—the questions that define a psychological theory’s explanatory domain are usually couched in intentional terms—but, as we have seen, it does not follow that the theory characterizes the states and processes it describes as necessarily intentional. Computational states and processes will typically have no true intentional description when considered independently of an environment. Intentional characterizations are therefore not part of the individuative apparatus of computational theories. In this sense, (C) is false. Yet the argument does contain an important insight: an intentional characterization is needed to connect a computational theory with its pretheoretic explananda. An explanation of how the visual system detects the depth of the scene from information contained in two-dimensional images is forthcoming only when the states characterized in formal terms by the theory are construed as representations of distal properties.

但是，有人可能会反驳说，计算过程的意向性解释所起的关键解释作用难道还不足以使计算理论具有意向性吗？事实上，当计算理论脱离了确保其解释相关性的意向性解释时，似乎不能正确地将其描述为认知理论。从某种意义上说，这是正确的；然而，这并不会破坏我的观点，即计算理论不是意向性的。让我解释一下。

But, one might object, isn’t this crucial explanatory role played by an intentional interpretation of a computational process enough to make the computational theory intentional? Indeed, it might seem that a computational theory, when divorced from the intentional interpretation that secures its explanatory relevance, cannot properly be characterized as a theory of cognition. There is a sense in which this is true; however, it does not undermine my point that computational theories are not intentional. Let me explain.

计算理论为机制计算的函数提供了数学特征，但只有在某些环境中，该函数才能被描述为认知函数（即，其参数和值在认识论上相关的函数，使得计算的输出在给定输入的情况下可以被视为合理的或令人信服的）。举个例子可以更清楚地说明这一点。根据 Marr 的说法，从双目视差计算深度所必需的立体图像匹配得益于我们世界的一个基本事实——视差变化平稳，因为物质具有凝聚力。这是 Marr 所说的自然约束（特别是连续性约束）的一个例子。在某些环境中，无法满足机制计算的数学函数的认知解释约束。在不满足连续性约束的环境中（尖峰宇宙），立体视觉模块将计算相同的正式特征函数，但不会从视差计算深度。该函数可能没有认知（即理性）在这个环境中描述。计算理论脱离实际环境，因为它旨在提供一种抽象的、因而是完全通用的机制描述，该描述为预测和解释其在任何环境中的行为提供了基础，即使在设备所做的事情不能被舒适地描述为认知的环境中也是如此。当计算特性伴随着适当的有意解释时，我们可以看到，在特定的环境中，计算特定数学函数的机制如何能够促进视觉等认知功能。

A computational theory provides a mathematical characterization of the function computed by a mechanism, but only in some environments can this function be characterized as a cognitive function (that is, a function whose arguments and values are epistemically related, such that the outputs of the computation can be seen as rational or cogent given the inputs). An example will make the point clearer. The matching of stereo images essential to the computation of depth from binocular disparity is aided, according to Marr, by a fundamental fact about our world—that disparity varies smoothly, because matter is cohesive. This is an example of what Marr calls a natural constraint (in particular, the continuity constraint). In some environments, the constraints that enable a cognitive interpretation of the mathematical function computed by a mechanism will not be satisfied. In environments where the continuity constraint is not satisfied (a spiky universe), the stereopsis module would compute the same formally characterized function, but it would not be computing depth from disparity. The function might have no cognitive (i.e., rational) description in this environment. A computational theory prescinds from the actual environment because it aims to provide an abstract, and hence completely general, description of a mechanism that affords a basis for predicting and explaining its behavior in any environment, even in environments where what the device is doing cannot comfortably be described as cognition. When the computational characterization is accompanied by an appropriate intentional interpretation, we can see how a mechanism that computes a particular mathematical function can, in a particular context, subserve a cognitive function such as vision.

计算理论通过将具有该能力的机制纳入抽象的计算描述来解释认知能力。通过参考具有独立理论特征的一类设备来解释理论前可识别的能力是一种在其他领域（尤其是生物学）中常见的解释策略。沙鲨探测猎物的能力是通过在鲨鱼体内假设一个电场探测器来解释的，该探测器是一种结构和行为以电磁理论为特征的设备。电磁理论在鲨鱼猎物探测能力的生物学解释中发挥了大部分解释作用。当然，这种解释也诉诸其他事实——例如，鲨鱼自然环境中的动物（而不是岩石和其他无生命物体）会产生显著的电场——但没有人会认为这些事实是电磁理论的一部分。同样，通过指定机制所属的计算设备类别并提供该类别行为的独立（即非认知）特征，计算理论在解释认知能力方面承担了主要的解释负担。对该过程的有意解释也起着解释作用——它表明能力已经得到解释——但在认知解释中发挥重要作用并不意味着它就成为计算理论本身的一部分。

A computational theory explains a cognitive capacity by subsuming the mechanism that has that capacity under an abstract computational description. Explaining a pretheoretically identifiable capacity by reference to a class of devices that have an independent, theoretical, characterization is an explanatory strategy familiar from other domains, particularly biology. The ability of sand sharks to detect prey is explained by positing within the shark the existence of an electric field detector a device whose architecture and behavior is characterized by electromagnetic theory. Electromagnetic theory does most of the explanatory work in the biological explanation of the shark’s prey detecting capacity. Of course, the explanation appeals to other facts—for example, that animals, but not rocks and other inanimate objects in the shark’s natural environment, produce significant electric fields—but no one would suggest that such facts are part of electromagnetic theory. Similarly, by specifying the class of computational devices to which a mechanism belongs and providing an independent (i.e., noncognitive) characterization of the behavior of this class, a computational theory bears the primary explanatory burden in the explanation of a cognitive capacity. The intentional interpretation of the process also plays an explanatory role—it demonstrates that the capacity has been explained—but playing an essential role in the cognitive explanation does not thereby make it part of the computational theory proper.

那么，计算理论不是认知理论吗？这要视情况而定。如果一个理论必须给出机制的认知特征（根据这一特征，计算认知功能是该机制的必要属性）才能成为认知理论，那么计算理论就不是认知理论。如果承担解释认知能力的主要解释负担就足够了，那么它们通常就是认知理论。

So does it follow that computational theories are not cognitive? It depends. If a theory must give a cognitive characterization of a mechanism (according to which computing a cognitive function is a necessary property of the mechanism) to be a cognitive theory, then computational theories are not cognitive. If bearing the primary explanatory burden in an explanation of a cognitive capacity is sufficient, then they typically are.

让我们简要回顾一下威尔逊的理论适当性条件，即科学解释在解释项和被解释项中以相同的水平描述现象。上述计算解释的说明表明，理论适当性并不是对科学解释的一般约束。计算解释以非故意、形式的方式描述认知能力。在任何情况下，这一要求都是不合理的，因为它不仅排除了先前表征现象的还原性解释（例如，微观还原），而且还排除了功能分析的解释，这是认知心理学和生物学中的主要解释形式。11 此类解释通常将^复杂的能力或过程分析为更基本的、不太专业的元素。例如，对进行长除法的能力的解释诉诸于复制数字和执行乘法和减法的能力。对消化的解释诉诸于更基本的化学过程。这两种解释似乎都违反了威尔逊的理论适当性条件。尽管威尔逊承认理论的适当性是对科学解释的一个可废止的限制，但这种解释的普遍性表明它根本不是一个限制。

Let us return briefly to Wilson’s theoretical appropriateness condition, the requirement that a scientific explanation characterize a phenomenon at the same level in the explanans as in the explanadum. The above account of computational explanation suggests that theoretical appropriateness is not a general constraint on scientific explanation. Computational explanations characterize cognitive capacities in nonintentional, formal, terms. The requirement is independently implausible in any case, since it would rule out not only reductive explanations (e.g., microreductions) of antecedently characterized phenomena, but also explanation by functional analysis, the predominant form of explanation in both cognitive psychology and biology.¹¹ Such explanations typically analyze complex capacities or processes into more basic, less specialized, elements. For example, the explanation of the capacity to do long division appeals to the ability to copy numerals and perform multiplication and subtraction. The explanation of digestion appeals to more basic chemical processes. Both of these explanations appear to violate Wilson’s theoretical appropriateness condition. Though Wilson grants that theoretical appropriateness is a defeasible constraint on scientific explanation, the ubiquity of explanations of this sort suggests that it is not a constraint at all.

14.4 内容的归属

14.4  The Ascription of Content

计算系统的解释由解释函数 f _I给出，该函数指定系统的假定结构与某个表示域的元素之间的映射。例如，要将设备解释为加法器，需要指定解释函数f _I ，该函数将设备的状态与数字配对。只有当存在一个解释函数，能够以相当直接的方式将形式表征的结构映射到这些元素时，才可以说该设备表示域中的元素。

An interpretation of a computational system is given by an interpretation function f_I that specifies a mapping between the postulated structures of the system and elements of some represented domain. For example, to interpret a device as an adder involves specifying an interpretation function f_I that pairs states of the device with numbers. The device can plausibly be said to represent elements in the domain only if there exists an interpretation function that maps formally characterized structures to these elements in a fairly direct way.

由于解释只是形式化特征元素与某个表示域元素之间的结构保留映射，因此没有理由认为计算系统的解释是唯一的。计算解释的非唯一性一直被认为是计算主义的问题，但事实上并非如此。大多数“非预期”解释都不符合直接性要求。12更重要的是，计算解释的合理性仅取决于是否存在具有解释作用的^解释。13

Since an interpretation is just a structure-preserving mapping between formally characterized elements and elements of some represented domain, there is no reason to think that the interpretation of a computational system will be unique. The non-uniqueness of computational interpretation has been thought to be a problem for computationalism, but in fact it is not. Most “unintended” interpretations will not meet the directness requirement.¹² More importantly, the plausibility of a computational account depends only on the existence of an interpretation that does explanatory work.¹³

如果上述对内容解释作用的说明是正确的，那么对计算系统的解释应该将理论的形式装置与其前理论的解释联系起来。这一要求将限制对适当解释函数的选择。一个旨在解释我们算术能力的计算理论，除非它所假设的一些状态可以解释为代表数字，否则它就不能合理地声称自己已经做到了这一点。该系统也可以被解释为绘制六日战争的进程（以乔治·雷伊为例），这一事实不会破坏理论家声称已经描述了一个算术系统的说法，假设该机制可以一致且直接地解释为计算适当的算术函数。鉴于上一节中描述的有意解释的解释作用，计算系统“非预期”解释的存在无关紧要。理论的预先存在的解释为内容的归属设定了条件。

If the above account of the explanatory role of content is correct, then the interpretation of a computational system should connect the formal apparatus of the theory with its pretheoretic explananda. This requirement will constrain the choice of an appropriate interpretation function. A computational theory that purports to explain our arithmetical abilities cannot plausibly claim to have done so unless some of the states it postulates are interpretable as representing numbers. The fact that the system could also be interpreted as charting the progress of the Six-Day War (to use an example of Georges Rey’s) would not undermine the theorist’s claim to have described an arithmetical system, assuming that the mechanism can be consistently and directly interpreted as computing the appropriate arithmetical functions. Given the explanatory role of intentional interpretation as characterized in the previous section, the existence of “unintended” interpretations of computational systems is irrelevant. The preexisting explananda of the theory set the terms for the ascription of content.

考虑一下这对于旨在解释我们感知能力的理论意味着什么。定义感知理论领域的认知任务通常是根据恢复有关主体正常环境的某些类型的信息来指定的。将系统状态解释为代表环境特定属性表明该理论解释了主体如何能够在其正常环境中恢复这些信息。因此，我们应该期望归因于计算特征感知状态的内容是广泛的，也就是说，在明显不同的环境中，物理上相同的主体不会共享这些内容。

Consider what this means for theories that purport to explain our perceptual capacities. The cognitive tasks that define the domains of theories of perception are typically specified in terms of the recovery of certain types of information about the subject’s normal environment. Interpreting states of the system as representing environment-specific properties demonstrates that the theory explains how the subject is able to recover this information in its normal environment. Consequently, we should expect the contents ascribed to computationally characterized perceptual states to be broad, that is, not shared by physically identical subjects in significantly different environments.

Fodor（例如 1980、1984、1987）和其他人（例如 Block，1986；和 Cummins，1989）认为，计算心理学必须将自己限制在狭义内容的概念上，即附加于主体内在物理状态的内容。14^这种观点的动机部分在于认识到计算分类学与主体的正常环境无关。物理重复就是计算重复。鉴于这一事实，如果计算状态本质上具有其语义属性，那么计算心理学就需要一种附加于系统物理属性的内容概念；换句话说，它需要一种狭义内容的概念。但是，如果正如我所说，计算状态仅非本质上具有其语义属性，那么狭义内容就没有必要。事实证明，计算心理学不应将自己限制在狭义内容上有充分的理由。

It has been argued by Fodor (e.g., 1980, 1984, 1987) and others (e.g., Block, 1986; and Cummins, 1989) that computational psychology must restrict itself to a notion of narrow content, that is, content that supervenes on intrinsic physical states of the subject.¹⁴ In part, the motivation for such a view is the recognition that computational taxonomy prescinds from the subject’s normal environment. Physical duplicates are computational duplicates. Given this fact, if computational states have their semantic properties essentially, then computational psychology requires a notion of content that supervenes on the physical properties of the system; in other words, it needs a notion of narrow content. But if, as I have argued, computational states have their semantic properties only nonessentially, then narrow content is not necessary. And it turns out that there are good reasons why computational psychology should not restrict itself to narrow content.

首先，狭义内容这一有用概念的指定难度众所周知。更重要的是，由于感知理论的解释项通常以特定于环境的术语来表述，因此普通的特定于环境的宽泛内容将最能满足此类理论的解释目标。这一点可以概括。人们普遍认为普通内容是宽泛的。只要计算理论的理论前解释项以普通术语来表达，将宽泛内容归因于计算状态和结构将是合适的。

In the first place, a useful notion of narrow content has been notoriously hard to specify. More importantly, since the explananda of theories of perception are typically formulated in environment-specific terms, ordinary environment-specific broad contents will best serve the explanatory goals of such theories. The point can be generalized. It is widely appreciated that ordinary contents are broad. Insofar as the pretheoretic explananda of computational theories are framed in ordinary terms, the ascription of broad content to computational states and structures will be appropriate.

仔细研究 Marr 的理论可以证实这一点。他尽可能地归因于广泛的、特定于环境的内容。如果在主体的正常环境中，结构与显着的远端属性可靠地相关，那么 Marr 会将该结构描述为代表该属性。（例如，他将 2.5D 草图中的结构描述为代表表面方向。）Marr 理论提出的一些结构与主体正常环境中没有简单的远端属性标记相关。Marr 称之为边缘的结构有时与表面方向的变化相关，有时与深度、照明或反射率的变化相关。Marr 将边缘描述为代表这种分离的远端属性。请注意，在这两种情况下 - 与主体正常环境中的简单远端属性相关或与主体正常环境中的分离远端属性相关 - 归因于表征结构的内容都是广泛的。此外，如此归因的广泛内容由主体正常环境中获得的相关性决定，而不是由其他环境中获得的相关性决定。

A close look at Marr’s theory confirms the point. He ascribes broad, environment-specific contents where possible. If in a subject’s normal environment a structure is reliably correlated with a salient distal property, then Marr describes the structure as representing that property. (For example, he describes structures in the 2.5D. sketch as representing surface orientation.) Some of the structures posited by Marr’s theory correlate with no simple distal property tokening in the subject’s normal environment. The structures that Marr calls edges sometimes correlate with changes in surface orientation, sometimes with changes in depth, illumination, or reflectance. Marr describes edges as representing this disjunctive distal property. Notice that in both cases—correlation with a simple distal property in the subject’s normal environment or correlation with a disjunctive distal property in the subject’s normal environment—the contents ascribed to the representational structures are broad. Moreover, the broad contents so ascribed are determined by the correlations that obtain in the subject’s normal environment, not by those that would obtain in some other environment.

然而，Marr 所假设的某些结构（例如，单个零交叉点）与主体正常环境中任何易于表征的远端属性（简单或分离）都不相关。它们的一些标记与远端属性相关，其他标记似乎只是成像过程的产物。Marr 认识到这一事实，并警告说，诸如零交叉点之类的结构并非“具有物理意义”；他将它们描述为表示图像中的不连续性。它们的内容只是近端的，因此是狭隘的——它们依赖于主体的内在属性。但这种近端或狭隘的内容远非 Marr 的选择内容，而是他的最后手段，因为只有在无法进行广泛的内容归属时，他才会归因于近端内容^。15

Some of the structures that Marr posits (e.g., individual zero-crossings) do not, however, correlate with any easily characterized distal property, simple or disjunctive, in the subject’s normal environment. Some of their tokenings correlate with distal properties, others appear to be mere artifacts of the imaging process. Marr recognizes this fact, cautioning that such structures as zero-crossings are not “physically meaningful”; he describes them as representing discontinuities in the image. Their contents are only proximal, and hence narrow—they supervene on the intrinsic properties of the subject. But such proximal or narrow content, far from being Marr’s content of choice, is his content of last resort, since he ascribes proximal content only when a broad content ascription is unavailable.¹⁵

内容的共变理论（或信息论理论）将表征状态的意义与特定情况下状态标记的原因联系起来。16^前述马尔理论中的内容归属说明可能会诱使一些人认为他的理论默认了内容的共变理论。这将是一个错误。我曾声称，在归因内容时，马尔会在主体的正常环境中寻找结构标记的显著远端相关物。我一直小心翼翼地避免声称这些相关物是结构标记的原因。虽然说它们是结构标记的原因可能很自然，但马尔并没有这样说，而且不这样做可以避免许多众所周知的问题。17应该清楚的是，如果考虑找不到结构标记的显著^远端相关物（简单或分离）的情况，马尔的理论并不致力于内容的共变理论。在这种情况下，Marr 会将近端内容归因于结构，将其解释为代表图像或输入表示的特征，而不是其标记的远端原因，无论它是什么。近端内容的归因具有重要的解释目的——它使设备的计算说明更加清晰，因为它允许我们跟踪设备在处理过程中所做的事情，而理论假设的结构与显着的远端属性没有完全关联。18^对这些结构进行不明确的远端解释不会有任何解释目的；因此，Marr 不会将它们解释为代表其远端原因。决定采用近端解释而不是远端解释完全是出于解释考虑^。19

Covariational (or information-theoretic) theories of content identify the meaning of a representational state with the cause of the state’s tokening in certain specifiable circumstances.¹⁶ The foregoing account of content ascription in Marr’s theory may tempt some to find in his theory a tacit endorsement of a covariational theory of content. This would be a mistake. I have claimed that in ascribing content Marr looks for salient distal correlates of a structure’s tokening in the subject’s normal environment. I have been careful to avoid claiming that these correlates are the cause of the structure’s tokening. Though it may be natural to say that they are, Marr makes no such claim, and a number of well-known problems are avoided by not doing so.¹⁷ It should be clear that Marr’s theory is not committed to a covariational theory of content if one considers the sort of case where no salient distal correlate (simple or disjunctive) of a structure’s tokening can be found. In such cases, Marr ascribes a proximal content to the structure, interpreting it as representing a feature of the image or input representation rather than the distal cause of its tokening, whatever that might be. The ascription of proximal content serves an important expository purpose—it makes the computational account of the device more perspicuous, by allowing us to keep track of what the device is doing at points in the processing where the theory posits structures that do not correlate neatly with a salient distal property.¹⁸ No explanatory purpose would be served by an unperspicuous distal interpretation of these structures; consequently, Marr does not interpret them as representing their distal causes. The decision to adopt a proximal rather than a distal interpretation is dictated by purely explanatory considerations.¹⁹

14.5 计算心理学和自然主义心理学

14.5  Computational Psychology and Naturalistic Psychology

计算描述和广泛内容的兼容性似乎在文献中被忽视了。Cummins (1989) 对计算心理学中的表征的解释与我的有些相似，他表示：

The compatibility of computational description and broad content seems to have gone unnoticed in the literature. Cummins (1989), whose account of representation in computational psychology bears some resemblance to mine, says the following:

卡明斯的错误在于，他认为计算理论试图提供认知过程的非意图、环境无关的特征，这意味着它无法解释用环境特定术语指定的现象。我们已经看到，这是错误的。计算理论通过将环境特定的认知能力纳入环境无关的特征来解释它。对过程的意图解释是理论提供的抽象特征与构成理论解释项的环境特定意图特征之间的桥梁。正因为意图解释在理论中并不发挥本质上的个体作用——​​换句话说，无论计算状态具有什么内容，它们都是非本质的——理论家可以自由地在适当的地方分配广泛的内容，以确保理论和解释项之间的联系。

Cummins’s mistake is in thinking that the fact that a computational theory seeks to provide a nonintentional, environment-independent characterization of a cognitive process entails that it cannot explain phenomena specified in environment-specific terms. This, we have seen, is wrong. A computational theory explains an environment-specific cognitive capacity by subsuming it under an environment-independent characterization. The intentional interpretation of the process serves as a bridge between the abstract characterization provided by the theory and the environment-specific intentional characterization that constitutes the theory’s explananda. Precisely because the intentional interpretation does not play an essentially individuative role in the theory—in other words, whatever contents computational states have, they have them non-essentially—the theorist is free to assign broad contents where appropriate to secure the connection between theory and explananda.

计算理论家能够并且通常会将广泛的内容分配给计算结构，这一事实对心理学具有更大的影响。在《方法论唯我论被视为认知心理学的一种研究策略》中，福多尔说：

The fact that the computational theorist can and typically will assign broad contents to computational structures has larger implications for psychology. In “Methodological Solipsism Considered as a Research Strategy in Cognitive Psychology,” Fodor says the following:

福多尔认为，自然主义心理学是有机体/环境关系的理论，它确定了我们心理术语的含义。他提出了两个论据来支持自然主义心理学不可能存在的说法。由于这两个论据在文献中都得到了彻底的阐述（参见福多尔 1980 年的评论），我在这里就不多说了。但据我所知，没有人质疑福多尔的暗示，即计算心理学和自然主义心理学是完全不相关的项目。如果我对内容在计算心理学中的作用的解释是正确的，那么福多尔的构想方式就是错误的。如果我们有一个完整的计算心理学，即对每个人类认知能力的计算解释，那么我们实际上就已经有一个自然主义心理学了。让我来详细阐述一下。

Naturalistic psychology, as Fodor construes it, is the theory of organism/environment relations that fix the meanings of our mental terms. He offers two arguments for the claim that a naturalistic psychology is impossible. As both arguments have been thoroughly worked over in the literature (see the commentaries that accompany Fodor, 1980), I won’t go into them here. But as far as I know, no one has disputed Fodor’s implication that computational psychology and naturalistic psychology are entirely unrelated projects. If my account of the role of content in computational psychology is correct, then Fodor’s way of conceiving things is wrong. If we had a complete computational psychology, that is, a computational account of each human cognitive capacity, we would ipso facto already have a naturalistic psychology. Let me elaborate.

尽管计算理论提供了与环境无关的正式过程特征，但理论家通常无法在不调查主体的正常环境的情况下发现正确的正式特征。通常，在指定认知机制计算的函数时，必要的第一步是发现使计算易于处理的环境约束。信息处理问题的解决方案通常由机制输入中包含的信息决定；解决方案只有在反映主体正常环境的非常普遍特征的附加信息的帮助下才能实现。例如，如前所述，从双眼视差计算深度是可能的，因为该机制的构建是为了假设其正常环境的真实情况——视差由于物质具有内聚性（连续性约束）而平稳变化。找到这种非常普遍的约束是描述机制必须解决的数学问题，从而得出正确的计算过程描述的必要第一步。

Although a computational theory provides a formal, environment-independent, characterization of a process, the theorist will usually be unable to discover the correct formal characterization without investigating the subject’s normal environment. Typically, a necessary first step in specifying the function computed by a cognitive mechanism is discovering environmental constraints that make the computation tractable. The solutions to information-processing problems are often underdetermined by information contained in the input to the mechanism; the solution is achieved only with the help of additional information reflecting very general features of the subject’s normal environment. For example, as previously mentioned, the computation of depth from binocular disparity is possible only because the mechanism is built to assume something that is true about its normal environment—that disparity varies smoothly because matter is cohesive (the continuity constraint). Finding constraints of this very general sort is a necessary first step in characterizing the mathematical problem that the mechanism has to solve, and thus in arriving at a correct computational description of the process.

计算理论家还有第二个更明显的作用，即对确定心理术语含义的有机体/环境相互作用做出具体说明，即对形式化特征化过程的有意解释做出具体说明。我曾指出，内容归属受到主体正常环境的限制。将内容归因于理论所假设的结构的过程涉及尝试指定这些结构标记的正常环境相关性。马尔成功地将远端内容归因于他的理论中假设的许多结构（马尔并不是唯一取得这一成就的人），这一事实表明自然主义心理学并非不可能。尽管计算心理学是形式化的——其分类学原理是形式化的——但它的发展与福多尔称之为自然主义心理学的项目密切相关^。20

There is a second and more obvious point at which the computational theorist will contribute to the specification of the organism/environment interactions that fix the meanings of mental terms—namely, in the specification of an intentional interpretation of a formally characterized process. I have argued that content ascription is constrained by the subject’s normal environment. The process of ascribing content to the structures posited by the theory involves the attempt to specify the normal environmental correlates of tokenings of these structures. The fact that Marr succeeded in ascribing distal contents to many of the structures posited in his theory (and Marr is not unique in this achievement) demonstrates that naturalistic psychology is not impossible. Although computational psychology is formal—its taxonomic principles are formal— it develops hand-in-glove with the project that Fodor calls naturalistic psychology.²⁰

14.6 账户范围和限制

14.6  Scope and Limits of the Account

我对内容解释作用的论述已通过参考经典计算模型进行了阐述和辩护。经典架构将认知过程视为规则控制的内部符号或数据结构的操作，这些符号或数据结构是明确的解释候选对象。相比之下，联结主义认知模型不假设定义设备操作的数据结构。（许多联结主义设备与经典设备不同，不能正确地描述为构建、存储和检索内部表示。）联结主义模型假设激活单元（节点）会增加或减少与其连接的其他单元的激活水平，直到整体稳定下来。因此，联结主义模型缺乏方便的“钩子”，无法将对过程的有意解释挂在其上。语义解释将内容归因于网络中的单个单元或一组单元的激活模式。但是，我认为上述对内容解释作用的论述没有理由不直接适用于联结主义系统。联结主义网络的语义解释与经典计算模型的解释一样，发挥着同样复杂的解释作用。最重要的是，它们将认知能力的联结主义理论与其前理论解释联系起来。

My account of the explanatory role of content has been articulated and defended by reference to classical computational models. Classical architectures treat cognitive processes as rule-governed manipulations of internal symbols or data structures that are explicit candidates for interpretation. Connectionist cognitive models, by contrast, do not posit data structures over which the device’s operations are defined. (Many connectionist devices, unlike classical devices, are not correctly described as constructing, storing, and retrieving internal representations.) Connectionist models posit activated units (nodes) that increase or decrease the level of activation of other units to which they are connected until the ensemble settles into a stable configuration. Consequently, connectionist models lack convenient “hooks” on which an intentional interpretation of a process may be hung. Semantic interpretations ascribe content either to individual units in the network or to patterns of activation over an ensemble of units. However, I see no reason why the above account of the explanatory role of content would not apply straightforwardly to connectionist systems. Semantic interpretations of connectionist networks play the same complex explanatory role as do interpretations of classical computational models. Most importantly, they connect a connectionist theory of a cognitive capacity with its pretheoretic explananda.

计算心理学是否会对意向状态（即信念和欲望）的典型案例提供很多启示，还有待观察。计算主义的显著成功在于表征高度模块化、信息封装的过程，例如早期视觉、句法和语音处理。这类理论所提出的状态未能表现出命题态度所特有的复杂功能作用（通常包括意识可及性）。它们是次信念状态。福多尔在《心灵的模块化》一书中对以正式的、计算的术语表征更为核心的认知过程（如信念固着）的前景表示了相当大的悲观态度（见本卷第 9 章）。我认为这种悲观态度是有道理的，因为信念归因的上下文敏感性使编程任务看起来难以处理。然而，如果命题态度的计算说明即将出现，内容将发挥与模块化能力理论相同的解释作用。这种可能性的一个有趣后果是，如此表征的命题态度本质上不会具有其内容。如果在相对不同的环境中标记类型相同的信念状态标记，则它们可能具有不同的内容。21因此，计算信念理论的前景挑战了正统心灵哲学的基本承诺。22^{有些人可能会得出结论，这种理论实际上与}命题态度无关，尽管民间对心灵的观念似乎没有任何内容可以证明这一结论。

It remains to be seen whether computational psychology will shed much light on paradigm cases of intentional states, namely, beliefs and desires. The conspicuous successes of computationalism have been in characterizing highly modularized, informationally encapsulated processes such as early vision and syntactic and phonological processing. The states posited by theories of this sort fail to exhibit the complex functional roles characteristic of the propositional attitudes (including, typically, accessibility to consciousness). They are subdoxastic states. Fodor, in The Modularity of Mind has expressed considerable pessimism about the prospects of characterizing in formal, computational terms more central cognitive processes such as belief fixation (see Chapter 9, this volume). I think this pessimism is well placed, if only because the context-sensitivity of belief ascription makes the programming task appear intractable. However, should a computational account of propositional attitudes be forthcoming, content would play the same explanatory role it plays in theories of modular capacities. An interesting consequence of this eventuality is that propositional attitudes, so characterized, would not have their contents essentially. Type-identical belief-state tokens might have different contents, should they be tokened in relevantly different environments.²¹ The prospect of a computational theory of belief, therefore, challenges a fundamental commitment of orthodox philosophy of mind.²² Some may conclude that such a theory would not really be about the propositional attitudes, though nothing in the folk conception of the mind would seem to warrant this conclusion.

第 14 章注释

Notes for Chapter 14

1.   1.对于类语言的表示，形式条件声称它们根据其语法具有因果角色。

2.   1.  For language-like representations, the formality condition claims that they have their causal roles in virtue of their syntax.

3.   2.有关Marr理论的最详细阐述，请参见（Marr，1982）。

4.   2.  For the most detailed exposition of Marr’s theory see (Marr, 1982).

5.   3. Loar (1988) 和 Segal (1989, 1991) 的观点可能最为接近。Loar 的提议关注常识，而非计算心理学。Segal 认为，狭义内容在 Marr 的理论中起着核心作用。有关对 Segal 提议的批评，请参阅 (Egan, 1996)。

6.   3.  Loar (1988) and Segal (1989, 1991) have perhaps come closest. Loar’s proposal concerns commonsense, as opposed to computational, psychology. Segal argues that narrow content plays a central role in Marr’s theory. See (Egan, 1996) for criticism of Segal’s proposal.

7.   4. Graves 等人 (1973) 的论证似乎暗含着对理论适当性的诉诸，他们认为，要解释说话者对其语言的了解，必须诉诸内化的语法知识。

8.   4.  A tacit appeal to theoretical appropriateness seems to underlie the argument of Graves et al. (1973) for the claim that the explanation of the speaker’s knowledge of her language must appeal to internalized knowledge of grammar.

9.   5. Shapiro (1993) 认为我的主张（在 Egan，1991 中）Marr 的理论不是有意的，这“令人吃惊”。它不应该令人吃惊。Field (1978) 的一个核心主张是，正如 Field 在其 (1986) 的论文中所说，“心理学理论有一个非意向性的核心”(114)。无论如何，Marr 的解释者并没有捍卫他的理论是有意的这一关键假设。

10.   5.  Shapiro (1993) has described my claim (in Egan, 1991) that Marr’s theory is not intentional as “startling.” It should not be startling. A central claim of Field (1978) is, as Field puts it in his (1986) paper, “that psychological theories have a non-intentional core” (114). In any event, interpreters of Marr have not defended the crucial assumption that his theory is intentional.

11.   6.在将马尔层次结构的各层级描述为描述层级时，我并不是要排除将按层级分类的项目视为现象和过程，而不是理论家可以使用的纯粹语言手段。

12.   6.  In describing the levels of Marr’s hierarchy as levels of description I do not mean to preclude treating the items classified by level as phenomena and processes rather than purely linguistic devices available to theorists.

13.   7.在论证计算理论的（各种）有意特征时，马尔的解释者指出，马尔将原始草图说成是“代表图像”，而其他结构代表诸如深度和表面反射率等远端特性。这些论点背后的假设是，马尔在这些段落中的话对于解决分类学问题具有决定性作用。如果理论解释如此简单，那么大部分科学哲学就会失效。科学理论的个体原则很少能从表达该理论的语言中读出。马尔的语言通常不谨慎或不一致。他没有理由这样做——他没有关注哲学家们关心的问题。然而，在我引用的文中段落中，马尔明确讨论了信息处理方法的基本承诺，特别是如何理解计算理论；因此这段话在当前问题的背景下具有特殊意义。

14.   7.  In arguing for (various) intentional characterizations of the theory of the computation, interpreters of Marr point out that he speaks of the primal sketch as “representing the image,” and of other structures as representing such distal properties as depth and surface reflectance. The assumption underlying such arguments is that Marr’s words in these passages are decisive for settling issues of taxonomy. If theory interpretation were so simple, much of the philosophy of science would be out of business. The individuative principles of a scientific theory can rarely be read off the language used to articulate the theory. Marr is not generally careful or consistent in his language. There is no reason why he should be—he is not focusing on the issues that have concerned philosophers. In the passage I have quoted in the text, however, Marr is explicitly discussing fundamental commitments of the information-processing approach, in particular, how the theory of the computation is to be understood; so the passage bears special significance in the context of the current issue.

15.   8. Colin McGinn 曾向我指出，计算理论在以下意义上是有意的：它确实指定了对计算过程的预期解释——预期解释是数学的。计算理论的最高层将系统描述为计算一系列基于数学实体定义的函数。我很高兴地说，计算理论在这种相当不寻常的意义上是有意的。这是 Marr 的解释者认为他的理论是有意的。（他们认为该理论本质上将系统描述为计算基于视觉领域各个方面定义的函数，而这正是我所否认的。）

16.   8.  Colin McGinn has pointed out to me that the theory of the computation is intentional in the following sense: it does specify an intended interpretation of a computational process—the intended interpretation is mathematical. The topmost level of a computational theory characterizes the system as computing a series of functions defined on mathematical entities. I am quite happy to say that a computational theory is intentional in this rather unusual sense. This is trot the sense in which interpreters of Marr have taken his theory to be intentional. (They have assumed that the theory characterizes the system, essentially, as computing a function defined on aspects of the visual domain, and this is precisely what I deny.)

17.   9 . 这并不是说理论家在试图制定设备的计算描述时可以忽略主体的环境。恰恰相反。参见第 14.5 节。

18.   9.  This is not to suggest that the theorist can ignore the subject’s environment in attempting to formulate a computational description of the device. Quite the contrary. See Section 14.5.

19. 10.这是 (Davies, 1991) 中一个例子的改编。

20. 10.  This is an adaption of an example from (Davies, 1991).

21. 11.有关泛函分析的说明，请参阅 Cummins (1983)，第 2 章。

22. 11.  See Cummins (1983), chap. 2, for an account of functional analysis.

23. 12.直接性要求排除了将桌子解释为加法器的可能性，因为将数字分配给桌子的状态需要解释器自己计算加法函数。系统没有做这项工作。直接性要求尚未明确规定，但请参阅 Cummins (1989)，第 8 章的讨论。我在这里略过这个问题，主要是因为，正如我们将在下面看到的，排除计算系统的非预期解释的“问题”通常不会出现。

24. 12.  The directness requirement precludes interpreting a desk as an adder, since the assignment of numbers to states of the desk requires the interpreter to compute the addition function herself. The system is not doing the work. The directness requirement has yet to be precisely specified, but see Cummins (1989), chap. 8 for discussion. I gloss over this issue here primarily because, as we shall see below, the “problem” of ruling out unintended interpretations of computational systems typically does not arise.

25. 13.存在多个满足直接性要求的解释仅仅表明，形式化特征化的设备能够计算多个认知功能。 上面描述的 visex 如果以不同的方式嵌入到生物体中，它将计算听觉领域的功能。

26. 13.  The existence of more than one interpretation meeting the directness requirement simply shows that the formally characterized device is capable of computing more than one cognitive function. The visex, described above, would compute a function on the auditory domain if it were embedded differently in the organism.

27. 14.其他人，例如 Stich (1983)，对内容归因通常与情境相关且与观察者相关的事实印象深刻，因此得出结论，认知心理学根本不应该关注内容。

28. 14.  Others, such as Stich (1983), impressed by the fact that content ascription is typically context-sensitive and observer-relative, have concluded that cognitive psychology should not advert to content at all.

29. 15 . 那些认为狭义内容是 Marr 所选择内容的评论者之所以这样做，大概是因为他们认识到，决定内容的相关性与远端属性在不同环境中可能会有很大差异（例如，参见 Segal，1989 年、1991 年）。他们没有注意到，对于 Marr 来说，相关相关性是主体在正常环境中获得的相关性。

30. 15.  Commentators who have thought narrow content to be Marr’s content of choice have presumably done so because they recognize that content-determining correlations with distal properties can vary wildly across environments (see, for example, Segal, 1989, 1991). They fail to notice that for Marr the relevant correlations are those that obtain in the subject’s normal environment.

31. 16 . 例如，参见 Stampe (1977)、Dretske (1981) 和 Fodor (1990)。当然，他们的论述存在重大差异。

32. 16.  See, for example, Stampe (1977), Dretske (1981), and Fodor (1990). There are, of course, important differences in their accounts.

33. 17.内容的共变理论的一个问题是，它们暗示符号的意义是由其所有潜在原因的分离而产生的。由于“马”标记不仅可能由马引起，也可能由看起来像马的牛引起，因此共变理论似乎暗示“马”是指马或看起来像马的牛。（有关讨论，请参阅 Fodor，1990。）“分离问题”引起了另一个困难，即，鉴于符号标记的所有潜在原因决定了其意义，如何解释符号错误表示其对象的可能性。虽然计算理论家对错误表示几乎没有什么可说的（他们关心的是描述在正常情况下发生的事情，即感知是真实的），但不难看出错误表示是如何根据我上面概述的内容归属而产生的。如果在标记的正常环境条件不满足的情况下对分配了远端内容（简单或分离）的结构进行标记，则它们将错误表示。假设作为军事训练演习的一部分，比尔（一个具有 Marrian 视觉系统的正常人）被安置在一个房间里，其中连续性约束（该约束认为由于物质具有凝聚力，视差会平稳变化）未得到满足。比尔的视觉系统通常根据视差信息计算深度。然而，在这种情况下，物质的尖峰向四面八方投射，比尔（或者更具体地说，比尔视觉系统的立体视觉模块）将计算与通常相同的形式表征函数，但他会将其他属性（不是我们有一个方便名称的属性）错误地表示为深度。一般来说，如果不满足通常使生物体能够计算认知功能的约束，它将无法表示其环境。

34. 17.  One problem with covariational theories of content is their implication that the meaning of a symbol is given by the disjunction of all of its potential causes. Since “horse” tokenings would be caused not only by horses, but also by horsey looking cows, covariational theories seem to imply that “horse” means horse or horsey looking cow. (See Fodor, 1990 for discussion.) The “disjunction problem” gives rise to a further difficulty, namely, how to account for the possibility of a symbol’s misrepresenting its object, given that all potential causes of a symbol’s tokening determine its meaning. Though computational theorists have had little to say about misrepresentation (their concern is to characterize what is going on in the normal case, where perception is veridical) it is not hard to see how misrepresentation can arise on the account of content ascription I have sketched above. Structures assigned distal contents (simple or disjunctive) will misrepresent if they are tokened when the normal environmental conditions for their tokening are not satisfied. Suppose that, as part of a military training exercise, Bill, a normal human with a Marrian visual system, is placed in a room where the continuity constraint, which holds that disparity varies smoothly because matter is cohesive, is not satisfied. Bill’s visual system normally computes depth from disparity information. However, in these circumstances, where spikes of matter project in all directions, Bill (or, more specifically, the stereopsis module of Bill’s visual system) will compute the same formally characterized function as he normally does, but he will misrepresent some other property (not a property for which we have a convenient name) as depth. In general, where the constraints that normally enable an organism to compute a cognitive function are not satisfied, it will fail to represent its environment.

35. 18.人们可能想知道我所说的“近端内容”是否真的是内容。可以肯定的是，近端内容与我们在普通的预测和解释实践中归因的内容没有太大相似之处；然而，我确实认为这种内容在内部过程的计算说明中发挥着真正的解释作用。再举一个例子，Marcus (1980) 将在自然语言理解过程中构建的结构描述解释为表示的不是远端对象（或公共语言句子），而是解析器堆栈或缓冲区中的项目。在视觉和解析情况下，将结构解释为表示在过程中较早构建的其他结构都具有重要的功能，使我们能够跟踪处理器正在做什么。鉴于计算心理学中内容归属的理由主要是解释性的，我认为近端内容应该被视为一种内容，尽管可能只是一种“最小”内容。

36. 18.  One might wonder whether what I am calling “proximal content” is really content at all. To be sure, proximal contents do not bear much resemblance to the contents we ascribe in our ordinary predictive and explanatory practices; however, I do think that contents of this sort play a genuine explanatory role in computational accounts of internal processes. To cite a second example, Marcus (1980) interprets the structural descriptions constructed in the course of natural language comprehension as representing not distal objects (or public language sentences) but the items in stacks or buffers of the parser. In both the vision and parsing cases, interpreting a structure as representing other structures constructed earlier in the process serves the important function of allowing us to keep track of what the processor is doing. Given that the rationale for content ascription in computational psychology is primarily explanatory, I think that proximal content should be treated as a species of content, though perhaps only as a sort of “minimal” content.

37. 19。Matthews (1988)、Segal (1989) 和 McGinn (1989) 指出了另一个反对对内容进行完全因果解释的理由。他们认为，心理表征的内容似乎部分取决于它们倾向于产生的行为类型。一个由裂缝和阴影共同引起的结构的标记是否意味着裂缝、阴影，还是裂缝或阴影，部分取决于其标记是否有助于产生适合裂缝、阴影或两者的行为。

38. 19.  Matthews (1988), Segal (1989), and McGinn (1989) note another reason to resist an exclusively causal account of content. They argue that the contents of mental representations seem to be partly determined by the sorts of behaviors that they tend to produce. Whether a structure whose tokening is caused by both cracks and shadows means crack, shadow, or crack or shadow depends in part upon whether its tokening contributes to the production of behavior appropriate to cracks, shadows, or both.

39. 20.自然主义心理学被解释为确定心理表征意义的有机体/环境关系的规范，不应与语义学中有时被称为“自然化项目”的事业相混淆。后者试图在非意图和非语义词汇中指定心理状态的意义的充分条件。这是一个纯粹的哲学项目，不是心理学家关心的问题。

40. 20.  Naturalistic psychology, construed as the specification of the organism/environment relations that fix the meanings of mental representations, should not be confused with the enterprise that is sometimes called “the naturalization project” in semantics. The latter attempts to specify sufficient conditions, in a nonintentional and nonsemantic vocabulary, for a mental state’s meaning what it does. It is a purely philosophical project, not the concern of psychologists.

41. 21 . 例如，计算信念理论会对我的水信念和我的孪生地球分身对水的信念进行类型识别，尽管适合我们各自世界的有意解释可能会为我们的类型相同的信念赋予不同的广泛内容。因此，计算信念理论会尊重作为狭义内容假设的主要动机的直觉——分身在心理相关方面是相同的，因此应该被纳入相同的心理概括之下。但由于计算理论并不致力于狭义内容，它也可以适应主体的环境是其信念内容的决定因素的直觉。

42. 21.  For example, a computational theory of belief would type-identify my water beliefs and my Twin Earth doppelganger’s twater beliefs, although intentional interpretations appropriate to our respective worlds might assign different broad contents to our type-identical beliefs. A computational theory of belief would, therefore, respect the intuition that has been the prime motivation for the postulation of narrow content—that doppelgangers are identical in psychologically relevant respects, and hence should be subsumed under the same psychological generalizations. But because a computational theory is not committed to narrow content, it can also accommodate the intuition that the subject’s environment is a determinant of her belief contents.

43. 22.但请参阅马修斯（Matthews，1994）对命题态度的解释，该解释否认命题态度本质上具有其内容。

44. 22.  But see Matthews (1994) for an account of propositional attitudes that denies that they have their contents essentially.

第四部分：建模世界

Part IV: Modeling the World

世界是复杂的。智能代理处理这种复杂性的一种方法是构建内部模型——简化的世界表示，同时仍保留有效行动所需的所有信息。本部分的论文介绍了各种建模方法，每种方法都侧重于表示不同类型的结构（分类结构、价值结构、预测结构和因果结构）。每种方法在理论上都卓有成效，既有助于理解人类的工作方式，也有助于构建在世界上有效行动的机器。

The world is complicated. One way that intelligent agents might deal with that complexity is to build internal models—simplified representations of the world that still preserve all of the information necessary for effective action. The papers in this part present a variety of modeling approaches, each focused on representing different kinds of structure (categorical structure, value structure, predictive structure, and causal structure). Each has been theoretically fruitful both for understanding how humans work and for building machines that act effectively in the world.

在第 15 章中，Cameron Buckner 讨论了从低级数据中提取有用抽象类别问题的经验主义方法的历史。他表明，现代深度神经网络可以看作是执行某些类型的转换，可以可靠地解决此问题。对神经网络及其功能的几何理解一直很重要；Buckner 展示了这一传统如何通过现代技术得以延续和扩展。

In chapter 15, Cameron Buckner discusses the history of empiricist approaches to the problem of extracting useful abstract categories from low-level data. He shows that modern deep neural networks can be seen as performing certain kinds of transformations that reliably solve this problem. The geometric understanding of neural networks and what they do has always been important; Buckner shows how this tradition is continued and expanded by modern techniques.

Julia Haas 的《评估思维》（第 16 章）讨论了强化学习模型。这些模型关注的是模拟世界以有效地在其中行动的代理，它们根据对其行动成功的反馈更新策略。Haas 的文章提供了一些见解，说明强化学习如何为在人工系统中构建欲望、价值和意志提供基础。她指出，在最具争议性的情况下，这最终导致了“奖励就足够”假设，该假设表明，所有智能行为最终都可以理解为试图最大化某些内在奖励信号。

Julia Haas’s “The Evaluative Mind” (chapter 16) discusses reinforcement learning models. These focus on agents who model the world to effectively act in it, updating policies in response to feedback about the success of their actions. Haas’s essay provides some insight into how reinforcement learning might provide a a basis for building desire, value, and the will into artificial systems. At its most controversial, she notes, this culminates in the ‘Reward Is Enough’ hypothesis, which suggests that all intelligent behavior can ultimately be understood in terms of trying to maximize some intrinsic reward signal.

在第 17 章中，摘录了安迪·克拉克 (Andy Clark) 的文章《下一步是什么？》中的一段话，他展示了预测编码模型的强大功能，该模型使用纠错反馈来更新世界模型并在其中有效地发挥作用。基于人工智能 (AI) 研究中不断出现的一个观点，克拉克认为大脑不需要构建一个完整的内部世界模型来完整准确地反映周围的世界；相反，大脑会学习如何在预测未来世界的相关状态时最小化错误信号。克拉克展示了这种建模视角的变化如何为一系列感知、认知和运动现象提供描述性和解释性杠杆，包括中枢神经系统疾病和各种幻觉。

In chapter 17, an excerpt from his article “Whatever Next?,” Andy Clark demonstrates the power of predictive coding models, which use error-correction feedback both to update models of the world and act effectively in it. Building on a constant refrain in critical artificial intelligence (AI) research, Clark argues that the brain has no need to construct a complete internal model of the world that reflects completely and accurately the world around it; instead, the brain learns how to minimize error signals in predicting relevant states of the world looking forward. Clark shows how this change of modeling perspective provides descriptive and explanatory leverage over a host of perceptual, cognitive, and motor phenomena, including disorders of the central nervous system and illusions of various sorts.

最后，Judea Pearl 的《机器学习的理论障碍》（第 18 章）讨论了最近在构建设备方面取得的进展，这些设备可以从相关数据以及更重要的关于干预对世界的影响的事实中学习其世界的因果结构。这样的模型允许代理了解其世界的显著特征，指导干预以产生预期结果，并诊断这些预测中的错误来源。Pearl 提出的那种模型现在通常用于理解复杂变量集中许多变量之间的因果关系，并已被用作人类发展的模型。在构建搜索和发现世界因果真相的机器时，人们为机器提供了构建世界因果结构内部表示的各种工具，这些知识可能被认为构成了我们对整个世界的大部分（或至少是相当大一部分）知识。

Finally, Judea Pearl’s “Theoretical Impediments to Machine Learning” (chapter 18) discusses recent advances in building devices that can learn the causal structure of their worlds from correlational data and, importantly, from facts about the consequences of interventions on the world. Such models allow agents to pick up on salient features of their world, guide interventions to produce desired outcomes, and diagnose the sources of error in those predictions. Models of the sort that Pearl advances are now routinely used to understand the causal relations among many variables in complex variable sets and have been used as models of human development. In building machines that search for and discover causal truths about the world, one gives machines the kinds of tools for building internal representations of the causal structure of the world, which knowledge might be thought to comprise the bulk (or at least a sizable portion) of our knowledge of the world as a whole.

一个问题（仍未得到解决）是，我们是否应该期望确定单一的建模策略，或者有效的心智设计是否需要不同领域的不同模型的组合。例如，强化学习和深度神经网络的结合是谷歌 AlphaGo 成功的关键。相反，预测编码的倡导者经常强调在感知和行动的每个阶段使用统一的建模策略所带来的理论和实践力量。正如我们所建议的，理解因果结构可能被认为是概念学习的先决条件（因为许多概念都涉及因果纠缠），预测接下来会发生什么，以及在复杂而混乱的世界中规划有效行动。

One question—still very much open—is whether we ought to expect to settle on a single modeling strategy or whether effective mind design requires a combination of different models in different domains. Combining reinforcement learning and deep neural networks, for example, was key to success for Google’s AlphaGo. Conversely, advocates of predictive coding often emphasize the theoretical and practical power that comes from using a unified modeling strategy for every stage of perception and action. And as we have suggested, understanding causal structure might be thought to be prerequisite for concept learning (as many concepts involve causal entanglements), for predicting what will happen next, and for planning effective actions in a complex and messy world.